diff --git a/backward/L0/100M/cfg.json b/backward/L0/100M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..481b05ab5e8c5d28de9186eb3048b303437b3bbf
--- /dev/null
+++ b/backward/L0/100M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.0.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 0, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L1", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1.0, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L0_hook_resid_post_L1_1_0", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/22ninmut", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/backward/L0/100M/sae_weights.safetensors b/backward/L0/100M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..aa456fad807f177cd3a80da4e4a025d2f66ec847
--- /dev/null
+++ b/backward/L0/100M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:eaf688bedd969130bc40dabef5ecd4d20d1650527d890da1486a3009cc52703e
+size 37801344
diff --git a/backward/L0/100M/sparsity.safetensors b/backward/L0/100M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..6a80499a2f05d43d21b14979046bce3847550b7b
--- /dev/null
+++ b/backward/L0/100M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3eccde1a921dbea545c2838d42de3e05112412e5ef0ff625013f1152fc54f41d
+size 24656
diff --git a/backward/L0/200M/cfg.json b/backward/L0/200M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..481b05ab5e8c5d28de9186eb3048b303437b3bbf
--- /dev/null
+++ b/backward/L0/200M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.0.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 0, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L1", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1.0, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L0_hook_resid_post_L1_1_0", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/22ninmut", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/backward/L0/200M/sae_weights.safetensors b/backward/L0/200M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..52492436b9b1189a55c00cd02dca8471ed62e6d4
--- /dev/null
+++ b/backward/L0/200M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b7e2e760369b076b49810ebb2e9856dbfb0cba2a555021dc7b1f4b181056d125
+size 37801344
diff --git a/backward/L0/200M/sparsity.safetensors b/backward/L0/200M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..4cdbaeebcb461eba5bfa95342bd2fbde0f0ca9f1
--- /dev/null
+++ b/backward/L0/200M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:174f1e22cce1b01e437592772e23c4c46d594bdd06361197c1f1d13c2de0adbe
+size 24656
diff --git a/backward/L0/300M/cfg.json b/backward/L0/300M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..481b05ab5e8c5d28de9186eb3048b303437b3bbf
--- /dev/null
+++ b/backward/L0/300M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.0.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 0, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L1", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1.0, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L0_hook_resid_post_L1_1_0", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/22ninmut", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/backward/L0/300M/sae_weights.safetensors b/backward/L0/300M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..3d69d2dde3bd09da95b104c22e16a7087efc656d
--- /dev/null
+++ b/backward/L0/300M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:90c6a2bc0fb1221f1a297e78f914dcd4431d0594d665b9e6971ad53ad1e3a4f9
+size 37801344
diff --git a/backward/L0/300M/sparsity.safetensors b/backward/L0/300M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..d63a442f19ab5dec0951ac8813348a08f41c2ce2
--- /dev/null
+++ b/backward/L0/300M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:82ab0163c40693019a4f9971942ad6b2bc7864077175efae7d2b48bc2772af5d
+size 24656
diff --git a/backward/L0/400M/cfg.json b/backward/L0/400M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..481b05ab5e8c5d28de9186eb3048b303437b3bbf
--- /dev/null
+++ b/backward/L0/400M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.0.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 0, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L1", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1.0, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L0_hook_resid_post_L1_1_0", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/22ninmut", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/backward/L0/400M/sae_weights.safetensors b/backward/L0/400M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..1179962c3a12d3e8cfe747aa314d2618ffd16f08
--- /dev/null
+++ b/backward/L0/400M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3523d121fcf5fba450df3a43ec1d64c3175d40673953b7525d740e02c61d1470
+size 37801344
diff --git a/backward/L0/400M/sparsity.safetensors b/backward/L0/400M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..a0c89388d6c4f398877f5b23a99a97141d4ffe41
--- /dev/null
+++ b/backward/L0/400M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ed8bdac9956cf39cd59d7407ca55c418010f28ef8e30c02221bcc3d2bac90e88
+size 24656
diff --git a/backward/L0/500M/.cfg.json.swp b/backward/L0/500M/.cfg.json.swp
new file mode 100644
index 0000000000000000000000000000000000000000..0ee340d20ffa189c1b170e84c0f2752c08db2560
Binary files /dev/null and b/backward/L0/500M/.cfg.json.swp differ
diff --git a/backward/L0/500M/cfg.json b/backward/L0/500M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..481b05ab5e8c5d28de9186eb3048b303437b3bbf
--- /dev/null
+++ b/backward/L0/500M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.0.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 0, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L1", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1.0, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L0_hook_resid_post_L1_1_0", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/22ninmut", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/backward/L0/500M/sae_weights.safetensors b/backward/L0/500M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..8ce597711368306cc3d966ec309aa82851c242b0
--- /dev/null
+++ b/backward/L0/500M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3f1f5172f0847c6a15fb993ea9c10fea00470bdfdb436eb9f5f3cd60c746e11d
+size 37801344
diff --git a/backward/L0/500M/sparsity.safetensors b/backward/L0/500M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..620847a5f98dcb4fa1c4247afb0d399d7d0ebb90
--- /dev/null
+++ b/backward/L0/500M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:53bb6ce0541567e2888b72807252ecd39b2d7763ce2dd9041f362a8913659437
+size 24656
diff --git a/backward/L1/100M/cfg.json b/backward/L1/100M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..30b421bfea79f4c8c16fa8d01c38528a56328a03
--- /dev/null
+++ b/backward/L1/100M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.1.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 1, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L2", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L1_hook_resid_post_L1_1", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/e902fx68", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/backward/L1/100M/sae_weights.safetensors b/backward/L1/100M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..91c594ab35c91cb1bdf76f41a210bc82984f4d45
--- /dev/null
+++ b/backward/L1/100M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:40654495475ccdfa37cba476e5365566e00c0b44cd848978c3a73c679570633b
+size 37801344
diff --git a/backward/L1/100M/sparsity.safetensors b/backward/L1/100M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..34c53d2068919b62572d719ba38c6d473cad23c0
--- /dev/null
+++ b/backward/L1/100M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a86592e4854a528529eab9693d810a1a5f2fea1e2a4dfebeffe23b1933caba07
+size 24656
diff --git a/backward/L1/200M/cfg.json b/backward/L1/200M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..30b421bfea79f4c8c16fa8d01c38528a56328a03
--- /dev/null
+++ b/backward/L1/200M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.1.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 1, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L2", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L1_hook_resid_post_L1_1", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/e902fx68", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/backward/L1/200M/sae_weights.safetensors b/backward/L1/200M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..d88ef81c78f96601ba4fa61ee32bc90da4683649
--- /dev/null
+++ b/backward/L1/200M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:faeff4d4e2cb069296723e02137a1f7c2262f85b9207a97ff0c44591a669b995
+size 37801344
diff --git a/backward/L1/200M/sparsity.safetensors b/backward/L1/200M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..1b8d9be6ae7ecade82eac9005425bccee9d12941
--- /dev/null
+++ b/backward/L1/200M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bbe8a880c3cd75eab5e4a2cff2565620fe824fb546a769b99cfac144975a7a80
+size 24656
diff --git a/backward/L1/300M/cfg.json b/backward/L1/300M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..30b421bfea79f4c8c16fa8d01c38528a56328a03
--- /dev/null
+++ b/backward/L1/300M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.1.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 1, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L2", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L1_hook_resid_post_L1_1", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/e902fx68", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/backward/L1/300M/sae_weights.safetensors b/backward/L1/300M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..dddbfe23d28f788d838b7cfa2a265d0fea2bcd7d
--- /dev/null
+++ b/backward/L1/300M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2512ab962c015caf7df3065b6f9029923b84a6605a9013d77b928741c4d88c2b
+size 37801344
diff --git a/backward/L1/300M/sparsity.safetensors b/backward/L1/300M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..4314cf3bc120eda3c25ab3dfd990f495e0f68092
--- /dev/null
+++ b/backward/L1/300M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e0bcbdff899ed867b0f7ad7541cc91b42157a2f4899454a7008ca22939ad7a94
+size 24656
diff --git a/backward/L1/400M/cfg.json b/backward/L1/400M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..30b421bfea79f4c8c16fa8d01c38528a56328a03
--- /dev/null
+++ b/backward/L1/400M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.1.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 1, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L2", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L1_hook_resid_post_L1_1", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/e902fx68", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/backward/L1/400M/sae_weights.safetensors b/backward/L1/400M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..60c2371f9afe7b5d672892f50829d9071ed70100
--- /dev/null
+++ b/backward/L1/400M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b83684c51c5d75b650c6d0c45216a426f53f1d8f29d11b612f12ba4221ac5f5b
+size 37801344
diff --git a/backward/L1/400M/sparsity.safetensors b/backward/L1/400M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..5d5242c3c38a536376fe7498da3ff1e5780516ac
--- /dev/null
+++ b/backward/L1/400M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4f505ab8c0d2a825afb14b676b4acef5c3c99827b93de3cba9990018cd6febc3
+size 24656
diff --git a/backward/L1/500M/cfg.json b/backward/L1/500M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..30b421bfea79f4c8c16fa8d01c38528a56328a03
--- /dev/null
+++ b/backward/L1/500M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.1.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 1, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L2", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L1_hook_resid_post_L1_1", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/e902fx68", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/backward/L1/500M/sae_weights.safetensors b/backward/L1/500M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..800593f595d2539f3c2336b7d2bb07c1dfdef838
--- /dev/null
+++ b/backward/L1/500M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:eede70447401ff9ab88d6e2a2d2bd52fae4035b2edbaca337bc67a9cd8ac2d03
+size 37801344
diff --git a/backward/L1/500M/sparsity.safetensors b/backward/L1/500M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..ac212c8c64725e7a7c2b4cfcfa0019d2517cc3de
--- /dev/null
+++ b/backward/L1/500M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c4b05feb5c4cf7cbb3ce4648e8fd36be3924fc33a9d942232ad189d1c460a267
+size 24656
diff --git a/backward/L10/100M/cfg.json b/backward/L10/100M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..ae54992ea0863c60482a816835eec98835f52c5e
--- /dev/null
+++ b/backward/L10/100M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.10.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 10, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L11", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L10_hook_resid_post_L1_1", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/7oj8g6c4", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/backward/L10/100M/sae_weights.safetensors b/backward/L10/100M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..9484257f06984efb4db5bbfd406553ff7449dd4a
--- /dev/null
+++ b/backward/L10/100M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:06c7f1afee5174577ebf48c4ba2b2623c6702f5efca1f78a234ad93fc2b430f9
+size 37801344
diff --git a/backward/L10/100M/sparsity.safetensors b/backward/L10/100M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..3af436766accbf713b34951f57fc548458c2926c
--- /dev/null
+++ b/backward/L10/100M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0df1c1f57067e205693519d17ea3ef8475d08e59d6fba88702ccaba5eba7e9de
+size 24656
diff --git a/backward/L10/200M/cfg.json b/backward/L10/200M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..ae54992ea0863c60482a816835eec98835f52c5e
--- /dev/null
+++ b/backward/L10/200M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.10.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 10, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L11", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L10_hook_resid_post_L1_1", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/7oj8g6c4", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/backward/L10/200M/sae_weights.safetensors b/backward/L10/200M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..9c4b77d183eb1435d158fea15bb4284312b4389c
--- /dev/null
+++ b/backward/L10/200M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2469b8f471cef975f3bb34af4032fe84028d291fe2e21caa5eb413c9decab866
+size 37801344
diff --git a/backward/L10/200M/sparsity.safetensors b/backward/L10/200M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..293930d39718bbf55bc5fb8ce013b6ba9fa43f01
--- /dev/null
+++ b/backward/L10/200M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9cf0a539cc13b22142a55193ddd9cf21e650016917f95265017723d483907c46
+size 24656
diff --git a/backward/L10/300M/cfg.json b/backward/L10/300M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..ae54992ea0863c60482a816835eec98835f52c5e
--- /dev/null
+++ b/backward/L10/300M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.10.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 10, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L11", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L10_hook_resid_post_L1_1", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/7oj8g6c4", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/backward/L10/300M/sae_weights.safetensors b/backward/L10/300M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..24b2412ed1a5b990ee07175df203f7283cb4361e
--- /dev/null
+++ b/backward/L10/300M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5108e31bc436352cee16984fd02e9694fee65469604e79827fd67c96e5577cbb
+size 37801344
diff --git a/backward/L10/300M/sparsity.safetensors b/backward/L10/300M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..f1364217c3d8217f609a87e39f9f6e61a1d63a7f
--- /dev/null
+++ b/backward/L10/300M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:564c97f8a8520504b79659fb02ee5ce0a8171ad865184fe7c61e17aedf6593dd
+size 24656
diff --git a/backward/L10/400M/cfg.json b/backward/L10/400M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..ae54992ea0863c60482a816835eec98835f52c5e
--- /dev/null
+++ b/backward/L10/400M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.10.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 10, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L11", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L10_hook_resid_post_L1_1", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/7oj8g6c4", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/backward/L10/400M/sae_weights.safetensors b/backward/L10/400M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..6f7e216ecda0fa9c112aabce98dc4c5ce6340edd
--- /dev/null
+++ b/backward/L10/400M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4f1faf430cb13561449cdd2c921a6f3f1c3208e786954644403c6c502629d630
+size 37801344
diff --git a/backward/L10/400M/sparsity.safetensors b/backward/L10/400M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..950dfcc73e173af4b2bb37325b9bbfc72008e201
--- /dev/null
+++ b/backward/L10/400M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1746ddb7c85e37d01196f381bf4c52cd3f4438e7814aa985205fadfc8bac887d
+size 24656
diff --git a/backward/L10/500M/cfg.json b/backward/L10/500M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..ae54992ea0863c60482a816835eec98835f52c5e
--- /dev/null
+++ b/backward/L10/500M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.10.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 10, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L11", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L10_hook_resid_post_L1_1", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/7oj8g6c4", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/backward/L10/500M/sae_weights.safetensors b/backward/L10/500M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..7a83eec2466cf02efce13d49b9584727c63d1e59
--- /dev/null
+++ b/backward/L10/500M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:46b26bfa7fd1c40589596c2635fb7b8f9f66733d37106eceef2ad086cee13357
+size 37801344
diff --git a/backward/L10/500M/sparsity.safetensors b/backward/L10/500M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..950605063bf04b0db75ae5f494a08e4fb7561375
--- /dev/null
+++ b/backward/L10/500M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:664f11b2ae8eb9e56008562107a2e88435b2985b96f83e0934c16005cf7dc7e8
+size 24656
diff --git a/backward/L2/100M/cfg.json b/backward/L2/100M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..80b9556319106e22d48b75195607da5316789e9a
--- /dev/null
+++ b/backward/L2/100M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.2.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 2, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L3", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L2_hook_resid_post_L1_1", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/is8g1rbr", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/backward/L2/100M/sae_weights.safetensors b/backward/L2/100M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..2e90af4c94ab825ca5725945970fec0078e1e459
--- /dev/null
+++ b/backward/L2/100M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8088d59e465ebf95e62b76e2fc5ac260da1a0baf3eb145ee7426bfc13ce609bb
+size 37801344
diff --git a/backward/L2/100M/sparsity.safetensors b/backward/L2/100M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..3e763c8c4740028620b3f0b4c385ab997291335d
--- /dev/null
+++ b/backward/L2/100M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b284c60af440f7662815ea431c073ae5c6e9c76685a4a447aed3e74b51ec1dd4
+size 24656
diff --git a/backward/L2/200M/cfg.json b/backward/L2/200M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..80b9556319106e22d48b75195607da5316789e9a
--- /dev/null
+++ b/backward/L2/200M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.2.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 2, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L3", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L2_hook_resid_post_L1_1", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/is8g1rbr", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/backward/L2/200M/sae_weights.safetensors b/backward/L2/200M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..4fcaf1db4e47439705f90f6d58ec8c9951c0abe5
--- /dev/null
+++ b/backward/L2/200M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:aeb0b7a2ded0e6dca599df75587d19a19b35d527013af2349bdd4e0c2e157116
+size 37801344
diff --git a/backward/L2/200M/sparsity.safetensors b/backward/L2/200M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..141b5495da89e0dfea5781b34188d4df58248028
--- /dev/null
+++ b/backward/L2/200M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:552111f6a4eeff0c5d2c6c4bf27aa252d48517852a3a7639bf84a6223076508c
+size 24656
diff --git a/backward/L2/300M/cfg.json b/backward/L2/300M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..80b9556319106e22d48b75195607da5316789e9a
--- /dev/null
+++ b/backward/L2/300M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.2.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 2, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L3", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L2_hook_resid_post_L1_1", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/is8g1rbr", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/backward/L2/300M/sae_weights.safetensors b/backward/L2/300M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..0a0ea4e8acf05ced64522f613f0ebec561b0dd85
--- /dev/null
+++ b/backward/L2/300M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a7f0156b7b0f70bd6b1fb43b075aa445b266465a289454cc86a0497b4dc7b60f
+size 37801344
diff --git a/backward/L2/300M/sparsity.safetensors b/backward/L2/300M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..20b1b2d404b0c0484f7fabe282a0b6eced6de407
--- /dev/null
+++ b/backward/L2/300M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0fb30f6789840fec89ee8a1567ae89845b8f145d418eda390360a70d6ca7adbd
+size 24656
diff --git a/backward/L2/400M/cfg.json b/backward/L2/400M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..80b9556319106e22d48b75195607da5316789e9a
--- /dev/null
+++ b/backward/L2/400M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.2.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 2, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L3", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L2_hook_resid_post_L1_1", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/is8g1rbr", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/backward/L2/400M/sae_weights.safetensors b/backward/L2/400M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..4f63d946caa7a82c2c5ce26f2934eec83a643e5c
--- /dev/null
+++ b/backward/L2/400M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:39ef3c0e2dcfe190f44c6cac7cac08f67afc47c24640648d41912b3db81f8190
+size 37801344
diff --git a/backward/L2/400M/sparsity.safetensors b/backward/L2/400M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..ee7a830d32c5dee3e6dfbb3c2181c1faebb5bdda
--- /dev/null
+++ b/backward/L2/400M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a2608e6ffcebb7d2f2e949ffe603780f9cabd5d2a659c2af08397483e2fe047b
+size 24656
diff --git a/backward/L2/500M/cfg.json b/backward/L2/500M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..80b9556319106e22d48b75195607da5316789e9a
--- /dev/null
+++ b/backward/L2/500M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.2.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 2, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L3", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L2_hook_resid_post_L1_1", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/is8g1rbr", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/backward/L2/500M/sae_weights.safetensors b/backward/L2/500M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..7815a87832f467509f39c2b24950430870576a91
--- /dev/null
+++ b/backward/L2/500M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:40b02004d239d83dd6dc70b0561fff4323ea85938483f4a6521f3eb7cb492ca1
+size 37801344
diff --git a/backward/L2/500M/sparsity.safetensors b/backward/L2/500M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..e86bfe903505bea0a09bd4b3864e5c45d7bc330e
--- /dev/null
+++ b/backward/L2/500M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1ba516ead1a1203d3dbda591a406fe1fbb989dc1b7f62f2ebd2c6eaf480b9a59
+size 24656
diff --git a/backward/L3/100M/cfg.json b/backward/L3/100M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..7479448673316736a22f2dde7cd0d4969b091f03
--- /dev/null
+++ b/backward/L3/100M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.3.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 3, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L4", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L3_hook_resid_post_L1_1", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/xxt7ihjo", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/backward/L3/100M/sae_weights.safetensors b/backward/L3/100M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..18bef9669a09b85081e7e18d837953b6c8a8e9b4
--- /dev/null
+++ b/backward/L3/100M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2825975c65965617026bf80ab4200705645a079cf76d0adfa65167e25f7d0b6a
+size 37801344
diff --git a/backward/L3/100M/sparsity.safetensors b/backward/L3/100M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..41c4ec8eea4bd149ccab23b21d4327e5daf8938f
--- /dev/null
+++ b/backward/L3/100M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:392022ac7bcf467f9cff8f49c048d4909423bacb04738c794d3d69a0d8656321
+size 24656
diff --git a/backward/L3/200M/cfg.json b/backward/L3/200M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..7479448673316736a22f2dde7cd0d4969b091f03
--- /dev/null
+++ b/backward/L3/200M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.3.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 3, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L4", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L3_hook_resid_post_L1_1", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/xxt7ihjo", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/backward/L3/200M/sae_weights.safetensors b/backward/L3/200M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..bd5c2560453aa317edda416c3a29645a9213aed7
--- /dev/null
+++ b/backward/L3/200M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e01edf4586fd5ff6394357717e2482a82e5d4811a71278fe8c079e00706f438c
+size 37801344
diff --git a/backward/L3/200M/sparsity.safetensors b/backward/L3/200M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..54492b98ec7ce16e38873efff660757443fe806d
--- /dev/null
+++ b/backward/L3/200M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0e4681efd0426d3559e3bcafb3bf4694a6e8844185af2afd11dc97a9800bde2d
+size 24656
diff --git a/backward/L3/300M/cfg.json b/backward/L3/300M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..7479448673316736a22f2dde7cd0d4969b091f03
--- /dev/null
+++ b/backward/L3/300M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.3.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 3, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L4", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L3_hook_resid_post_L1_1", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/xxt7ihjo", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/backward/L3/300M/sae_weights.safetensors b/backward/L3/300M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..f7cf8e36f6d2f2064dab7fb654559cbf753c997b
--- /dev/null
+++ b/backward/L3/300M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3d14ef1a0159aa1a99a8a91626918d6e025b33f79f5ee468ea811cd01dd4978c
+size 37801344
diff --git a/backward/L3/300M/sparsity.safetensors b/backward/L3/300M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..c7c4a5d22809fe4f9921f34c2ffd38577d4ea138
--- /dev/null
+++ b/backward/L3/300M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8cafd2fe492d791fe5a7576b267224fc941fead3049b087191e786ce42913716
+size 24656
diff --git a/backward/L3/400M/cfg.json b/backward/L3/400M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..7479448673316736a22f2dde7cd0d4969b091f03
--- /dev/null
+++ b/backward/L3/400M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.3.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 3, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L4", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L3_hook_resid_post_L1_1", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/xxt7ihjo", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/backward/L3/400M/sae_weights.safetensors b/backward/L3/400M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..8d7c975f72df38cd154969598206a1ba5c812837
--- /dev/null
+++ b/backward/L3/400M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:aac87ed542a6f4925cb0135ea8eb2f21811b110e4fabbff5feb5f656260142eb
+size 37801344
diff --git a/backward/L3/400M/sparsity.safetensors b/backward/L3/400M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..5b1d5bb9ae6171be5931ba55192db09d3cc24fb2
--- /dev/null
+++ b/backward/L3/400M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0bd88845ec317d32e113286382ccab09a8e5ff437f98d100f39937331d110b9b
+size 24656
diff --git a/backward/L3/500M/cfg.json b/backward/L3/500M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..7479448673316736a22f2dde7cd0d4969b091f03
--- /dev/null
+++ b/backward/L3/500M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.3.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 3, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L4", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L3_hook_resid_post_L1_1", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/xxt7ihjo", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/backward/L3/500M/sae_weights.safetensors b/backward/L3/500M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..d66e1f05991d86ca4572cb5c182f2b266ee15f7f
--- /dev/null
+++ b/backward/L3/500M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f1b831228b16ead1632c4ddcb9db0713d555bde31c1cf07387d89995f3293f9f
+size 37801344
diff --git a/backward/L3/500M/sparsity.safetensors b/backward/L3/500M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..a706168d2cc87834c52549db7a1ae80d6a561226
--- /dev/null
+++ b/backward/L3/500M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1a81e22d9080510bf76fa6ff0572e87b8f59144f7fa6537cedc7ea00e31d70c7
+size 24656
diff --git a/backward/L4/100M/cfg.json b/backward/L4/100M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..569aa3c63bb70f121421c85d102da99aa35e5355
--- /dev/null
+++ b/backward/L4/100M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.4.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 4, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L5", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L4_hook_resid_post_L1_1", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/h76le4wl", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/backward/L4/100M/sae_weights.safetensors b/backward/L4/100M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..68f91dbe45eea38192339d4d53e529d410361df9
--- /dev/null
+++ b/backward/L4/100M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:37438cb656456f218cac9afe49ffc2890d387c946d79cff4ad4284d74fa4c0dc
+size 37801344
diff --git a/backward/L4/100M/sparsity.safetensors b/backward/L4/100M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..63bd563184dfcf85101193ea2571f57ac12f1597
--- /dev/null
+++ b/backward/L4/100M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:231a8e5084b1deb2f21129fa81166d9a11e15316f6003167ffceeb8a1e9b7837
+size 24656
diff --git a/backward/L4/200M/cfg.json b/backward/L4/200M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..569aa3c63bb70f121421c85d102da99aa35e5355
--- /dev/null
+++ b/backward/L4/200M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.4.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 4, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L5", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L4_hook_resid_post_L1_1", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/h76le4wl", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/backward/L4/200M/sae_weights.safetensors b/backward/L4/200M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..a5eb0b65cbdb17f9ed80790a547990e6b44a5967
--- /dev/null
+++ b/backward/L4/200M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7e4b0c2a22db2b4dc2cd66823d54a888163dcb9dc99c67a589188b7aa250ce64
+size 37801344
diff --git a/backward/L4/200M/sparsity.safetensors b/backward/L4/200M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..544086d020d288c60d36d1ed215c47eae09d54d2
--- /dev/null
+++ b/backward/L4/200M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b415d6ac03136f83f302f34bfe0a35c1dd375be36bce71755a8588ba4328fc2f
+size 24656
diff --git a/backward/L4/300M/cfg.json b/backward/L4/300M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..569aa3c63bb70f121421c85d102da99aa35e5355
--- /dev/null
+++ b/backward/L4/300M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.4.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 4, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L5", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L4_hook_resid_post_L1_1", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/h76le4wl", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/backward/L4/300M/sae_weights.safetensors b/backward/L4/300M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..2d34a7dc9f1d707bb5cf846a89460e77acc54c02
--- /dev/null
+++ b/backward/L4/300M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:663f7e3f04d5263084a7efdc57d755694259c949262d3c8ed4cd088e4b0a2af6
+size 37801344
diff --git a/backward/L4/300M/sparsity.safetensors b/backward/L4/300M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..0ae3fff47b84b6993ec4195cb1cf71d6cedcc425
--- /dev/null
+++ b/backward/L4/300M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a6dd871fb603b12805948c6da3b01a980c58d8c4b2539750cd3588e7f1bc9882
+size 24656
diff --git a/backward/L4/400M/cfg.json b/backward/L4/400M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..569aa3c63bb70f121421c85d102da99aa35e5355
--- /dev/null
+++ b/backward/L4/400M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.4.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 4, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L5", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L4_hook_resid_post_L1_1", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/h76le4wl", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/backward/L4/400M/sae_weights.safetensors b/backward/L4/400M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..4e1688c5906a674af51ddb5f93cb9c024452b0ae
--- /dev/null
+++ b/backward/L4/400M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ac2fc030444c3233c6db0d45217e1cc0af263efd3f0e9bd4a0d34e3c69181d1d
+size 37801344
diff --git a/backward/L4/400M/sparsity.safetensors b/backward/L4/400M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..288de2b179af9ee11af797a16b57c96d2f2be6f1
--- /dev/null
+++ b/backward/L4/400M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:eb0339115daa9f799994bc0920fb7915be8161329e4e691191ad597ac655b9e0
+size 24656
diff --git a/backward/L4/500M/cfg.json b/backward/L4/500M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..569aa3c63bb70f121421c85d102da99aa35e5355
--- /dev/null
+++ b/backward/L4/500M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.4.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 4, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L5", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L4_hook_resid_post_L1_1", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/h76le4wl", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/backward/L4/500M/sae_weights.safetensors b/backward/L4/500M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..a7a700c936a445e8f212f70d4e5e9219c3f7de17
--- /dev/null
+++ b/backward/L4/500M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2f0534b7cae2b12ad21540aa8a88ca71322e0e0e94a25f4338eb2209ffd41e15
+size 37801344
diff --git a/backward/L4/500M/sparsity.safetensors b/backward/L4/500M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..7ce3fad90a67a180465232bd752cfdf15bbb00e3
--- /dev/null
+++ b/backward/L4/500M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:15d893250c9b29543fd70c52e5d9fd50c172e4d3fc62ed856e2c6ea24f1b1740
+size 24656
diff --git a/backward/L5/100M/cfg.json b/backward/L5/100M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..1e941b03012790f7ceb0322fdea8f31f3ab95162
--- /dev/null
+++ b/backward/L5/100M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.5.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 5, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L6", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L5_hook_resid_post_L1_1", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/jahbro6e", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/backward/L5/100M/sae_weights.safetensors b/backward/L5/100M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..9a9ca1d7c05d1e3920f6d3169ec2aedfaf091501
--- /dev/null
+++ b/backward/L5/100M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d0768b87e9e0169820ed3aa5d36bf528df27a69eeeb147a0405f0b88f03e9da8
+size 37801344
diff --git a/backward/L5/100M/sparsity.safetensors b/backward/L5/100M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..39bc22f358475e7d6d2df2d4028cb0783f4877b4
--- /dev/null
+++ b/backward/L5/100M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:71e3e4765510ec60f7142c773351bf1b02e993f60287c5ea5af2229484d0f334
+size 24656
diff --git a/backward/L5/200M/cfg.json b/backward/L5/200M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..1e941b03012790f7ceb0322fdea8f31f3ab95162
--- /dev/null
+++ b/backward/L5/200M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.5.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 5, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L6", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L5_hook_resid_post_L1_1", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/jahbro6e", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/backward/L5/200M/sae_weights.safetensors b/backward/L5/200M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..3e70c4d8da967eb3f78d2c8802fb41b4fef2f5f5
--- /dev/null
+++ b/backward/L5/200M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a106c95d733a25e4c1420424d9cea445e675805437731b8820dfb946a62d96cc
+size 37801344
diff --git a/backward/L5/200M/sparsity.safetensors b/backward/L5/200M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..a21ebdb299f996afe553be1a23e75ddb4f034a3c
--- /dev/null
+++ b/backward/L5/200M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3df10714787ae6411b0c50156091acfb8b966cbe6e7d75ba7a3664e9f4d26a0f
+size 24656
diff --git a/backward/L5/300M/cfg.json b/backward/L5/300M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..1e941b03012790f7ceb0322fdea8f31f3ab95162
--- /dev/null
+++ b/backward/L5/300M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.5.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 5, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L6", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L5_hook_resid_post_L1_1", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/jahbro6e", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/backward/L5/300M/sae_weights.safetensors b/backward/L5/300M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..da2ed959c87f0eeee127ec7380e195dc170f7096
--- /dev/null
+++ b/backward/L5/300M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c2c7d23f15c6c8269d58ee14906ef5cc62a0d201cbd723158d7dd53edc62afbf
+size 37801344
diff --git a/backward/L5/300M/sparsity.safetensors b/backward/L5/300M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..18eda007c5c7ee8cc6245875893b4e1221657e7f
--- /dev/null
+++ b/backward/L5/300M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dd38c5f9784ca7fc3452adc3ae6919d0556ced0320c35e62f2fc9b5868624c11
+size 24656
diff --git a/backward/L5/400M/cfg.json b/backward/L5/400M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..1e941b03012790f7ceb0322fdea8f31f3ab95162
--- /dev/null
+++ b/backward/L5/400M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.5.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 5, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L6", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L5_hook_resid_post_L1_1", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/jahbro6e", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/backward/L5/400M/sae_weights.safetensors b/backward/L5/400M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..0be54b5c3e2c634bdb90d6e9ed11761406233192
--- /dev/null
+++ b/backward/L5/400M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:517fa74aa700fcf6ed709a29838e8c4de75fc866692a1619d97135939c42e71b
+size 37801344
diff --git a/backward/L5/400M/sparsity.safetensors b/backward/L5/400M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..07e0f797749c2d14e81c6424c4a1df1bd6faccd6
--- /dev/null
+++ b/backward/L5/400M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6f0c2be80da4d64cb4db00aac9d986b18ea3ad1e98b6de09f66bc7f7d561cc83
+size 24656
diff --git a/backward/L5/500M/cfg.json b/backward/L5/500M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..1e941b03012790f7ceb0322fdea8f31f3ab95162
--- /dev/null
+++ b/backward/L5/500M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.5.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 5, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L6", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L5_hook_resid_post_L1_1", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/jahbro6e", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/backward/L5/500M/sae_weights.safetensors b/backward/L5/500M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..a4d24a6796e258b55e77524824b5f316b414c309
--- /dev/null
+++ b/backward/L5/500M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2d6b03a4472d7cfb95f2790af10c44ce3f6d6ccc45f9815323ac249ce2711898
+size 37801344
diff --git a/backward/L5/500M/sparsity.safetensors b/backward/L5/500M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..a03ec52261f90ddb9806a376e0f339960612e139
--- /dev/null
+++ b/backward/L5/500M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:93031963a7e0d355ded5944313e4d36dcea9c5749d8b8ff605e032101a18a9ad
+size 24656
diff --git a/backward/L6/100M/cfg.json b/backward/L6/100M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..fc8bc157fe3867d28d367eb338c31afa2e04b97e
--- /dev/null
+++ b/backward/L6/100M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.6.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 6, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L7", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L6_hook_resid_post_L1_1", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/epfkz598", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/backward/L6/100M/sae_weights.safetensors b/backward/L6/100M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..00c5fca8b6628a5185d2400f391f3ce39e885762
--- /dev/null
+++ b/backward/L6/100M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c43c06472003fa8863482a4d31179ba66e2d11bcb4c9b494ed68132289cdd9d4
+size 37801344
diff --git a/backward/L6/100M/sparsity.safetensors b/backward/L6/100M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..84ca831ad23798defd31f740a4bd45c2d1768209
--- /dev/null
+++ b/backward/L6/100M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:22c53cd0b0303427597edf6df858f9d619225fed1f38ebec2ff064cfb23e344e
+size 24656
diff --git a/backward/L6/200M/cfg.json b/backward/L6/200M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..fc8bc157fe3867d28d367eb338c31afa2e04b97e
--- /dev/null
+++ b/backward/L6/200M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.6.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 6, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L7", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L6_hook_resid_post_L1_1", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/epfkz598", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/backward/L6/200M/sae_weights.safetensors b/backward/L6/200M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..abae81f9ee6fe13b533fd386858868989ba3ad70
--- /dev/null
+++ b/backward/L6/200M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6a88dbced960067b27e7c48b5a35d6d8e80261211f95f7aa0a1d3fa29a0e3920
+size 37801344
diff --git a/backward/L6/200M/sparsity.safetensors b/backward/L6/200M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..fdae5dd2dbb7a2525023fd42d159243c9ef2f86d
--- /dev/null
+++ b/backward/L6/200M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b9170f641e430620213573f6298e02af004b5442ff6b69a747c71fddcee70ec8
+size 24656
diff --git a/backward/L6/300M/cfg.json b/backward/L6/300M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..fc8bc157fe3867d28d367eb338c31afa2e04b97e
--- /dev/null
+++ b/backward/L6/300M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.6.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 6, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L7", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L6_hook_resid_post_L1_1", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/epfkz598", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/backward/L6/300M/sae_weights.safetensors b/backward/L6/300M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..ac4eec132aa96af8f7d8986ab6631a827d17accb
--- /dev/null
+++ b/backward/L6/300M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a200030381c7771e39d384bcb4ec2aad8c48113d21ca1ce2753eb5a03ecdc525
+size 37801344
diff --git a/backward/L6/300M/sparsity.safetensors b/backward/L6/300M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..4be424796f4a304c2ffa2a98d82930e0cad7bdc3
--- /dev/null
+++ b/backward/L6/300M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d16b8c33ff98e1077ac2507f0f3ae05039d3d2ce2c370a81b1cf6e18a5e9e25a
+size 24656
diff --git a/backward/L6/400M/cfg.json b/backward/L6/400M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..fc8bc157fe3867d28d367eb338c31afa2e04b97e
--- /dev/null
+++ b/backward/L6/400M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.6.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 6, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L7", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L6_hook_resid_post_L1_1", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/epfkz598", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/backward/L6/400M/sae_weights.safetensors b/backward/L6/400M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..9b2b2d6a13743bf75da5a946fe7c88d30c58a709
--- /dev/null
+++ b/backward/L6/400M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9904604b5a15528f04d79289199beab899f85ac23716c0faa7744e38ceb02c62
+size 37801344
diff --git a/backward/L6/400M/sparsity.safetensors b/backward/L6/400M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..44ee79592f98084bb40aa09fb21c10c60bc55819
--- /dev/null
+++ b/backward/L6/400M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b870949450992706701a2ae6e39b74c2ba2205bcda46ef1c90f67eb1f1ac62ba
+size 24656
diff --git a/backward/L6/500M/cfg.json b/backward/L6/500M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..fc8bc157fe3867d28d367eb338c31afa2e04b97e
--- /dev/null
+++ b/backward/L6/500M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.6.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 6, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L7", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L6_hook_resid_post_L1_1", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/epfkz598", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/backward/L6/500M/sae_weights.safetensors b/backward/L6/500M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..446fae9e2d4a209a06b63a4bbd903fff45aa9284
--- /dev/null
+++ b/backward/L6/500M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:acae715ffa216bd94c800194c257ebe51b41f55795ff6bb0f5081fca7d712a26
+size 37801344
diff --git a/backward/L6/500M/sparsity.safetensors b/backward/L6/500M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..08311ca31800fe6e02e16ea7494633b4fcd4b42a
--- /dev/null
+++ b/backward/L6/500M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1be103810fc424d00209ecf4cb653a8bf06dae505e1d1d17433ca2ac40d3bdca
+size 24656
diff --git a/backward/L7/100M/cfg.json b/backward/L7/100M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..cc5047640a2fb1ee5f7ecf4cfadb049c99cfadc1
--- /dev/null
+++ b/backward/L7/100M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.7.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 7, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L8", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L7_hook_resid_post_L1_1", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/efgybgxl", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/backward/L7/100M/sae_weights.safetensors b/backward/L7/100M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..fad81de676c184a0868d694e3197e04c912cf7e8
--- /dev/null
+++ b/backward/L7/100M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7a6b7bb089ad0220539a66c83213f3ccb57a880e918764bf106a6ac1a780b9c5
+size 37801344
diff --git a/backward/L7/100M/sparsity.safetensors b/backward/L7/100M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..4b884640719cb82b5404d8e44414277f5b5512a5
--- /dev/null
+++ b/backward/L7/100M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ef2e3b483f5d038ae621b0b550c350ee17cc00820d4c1af004265670fa84d28a
+size 24656
diff --git a/backward/L7/200M/cfg.json b/backward/L7/200M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..cc5047640a2fb1ee5f7ecf4cfadb049c99cfadc1
--- /dev/null
+++ b/backward/L7/200M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.7.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 7, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L8", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L7_hook_resid_post_L1_1", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/efgybgxl", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/backward/L7/200M/sae_weights.safetensors b/backward/L7/200M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..6adbbc69bd7b6301cc35f3cb53152f0553b2b618
--- /dev/null
+++ b/backward/L7/200M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:772ba6e9b9826ebb9fadc0cc8304f53921da1b2b7dc86c49ec82f217963d75c2
+size 37801344
diff --git a/backward/L7/200M/sparsity.safetensors b/backward/L7/200M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..5d659b221b150f551a37368defcf1dfbca58896a
--- /dev/null
+++ b/backward/L7/200M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a018b42b4c0ec0b2a034b0e486074c185ec34feb952a895c498cabfbcf1cfbf4
+size 24656
diff --git a/backward/L7/300M/cfg.json b/backward/L7/300M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..cc5047640a2fb1ee5f7ecf4cfadb049c99cfadc1
--- /dev/null
+++ b/backward/L7/300M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.7.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 7, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L8", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L7_hook_resid_post_L1_1", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/efgybgxl", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/backward/L7/300M/sae_weights.safetensors b/backward/L7/300M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..1278ab9b1aeb9c82dbbaab48f8b23e73d03553ce
--- /dev/null
+++ b/backward/L7/300M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:95d9fd138ecd69456ed2579e8222424e14b5bae788f03e21a77b3daa96a5ab64
+size 37801344
diff --git a/backward/L7/300M/sparsity.safetensors b/backward/L7/300M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..55cde25eed5c2c73045dc9fdad1b9a5e8fa4b6a1
--- /dev/null
+++ b/backward/L7/300M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:91e07f91e4d53e477a0790b5ea56806ab746e970d4581414bf66f95c5949dd04
+size 24656
diff --git a/backward/L7/400M/cfg.json b/backward/L7/400M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..cc5047640a2fb1ee5f7ecf4cfadb049c99cfadc1
--- /dev/null
+++ b/backward/L7/400M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.7.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 7, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L8", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L7_hook_resid_post_L1_1", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/efgybgxl", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/backward/L7/400M/sae_weights.safetensors b/backward/L7/400M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..e5e57590184fc2198c6f890757e9723ab873ad05
--- /dev/null
+++ b/backward/L7/400M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b7b3ad77c90c8ccad53ee1f796dad7e74f8461fbdb173adbfda7cbc752b683ca
+size 37801344
diff --git a/backward/L7/400M/sparsity.safetensors b/backward/L7/400M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..6d212c5b03ec570485a36ddb0b3f91c3373de199
--- /dev/null
+++ b/backward/L7/400M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9bebfa50f11e49d36f7bbbc30dbe0f7c082c49c61cc1b8bb016651a5d1cd59a3
+size 24656
diff --git a/backward/L7/500M/cfg.json b/backward/L7/500M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..cc5047640a2fb1ee5f7ecf4cfadb049c99cfadc1
--- /dev/null
+++ b/backward/L7/500M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.7.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 7, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L8", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L7_hook_resid_post_L1_1", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/efgybgxl", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/backward/L7/500M/sae_weights.safetensors b/backward/L7/500M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..0b747880eb40050468300b9ae9d344447bed0a7e
--- /dev/null
+++ b/backward/L7/500M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c76898c39430ce32559411bccaedb6992eea8f1bd6b7304a8f2a178777079088
+size 37801344
diff --git a/backward/L7/500M/sparsity.safetensors b/backward/L7/500M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..2cda4f0b3e8b0e0149bb08dcc8aa4accff100c75
--- /dev/null
+++ b/backward/L7/500M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:63070e572392533cfe3e2323ac0dd0a52c6c50fadf771625a6a7a6f260120ee6
+size 24656
diff --git a/backward/L8/100M/cfg.json b/backward/L8/100M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..7a31d30ec1d468bdacdb4b32d12c742f8842e980
--- /dev/null
+++ b/backward/L8/100M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.8.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 8, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L9", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L8_hook_resid_post_L1_1", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/djw9vsut", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/backward/L8/100M/sae_weights.safetensors b/backward/L8/100M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..640bfea3ac5b64e63907341e50ec7ec081670b02
--- /dev/null
+++ b/backward/L8/100M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c88a94f94fa76d2443334202b792bcb2b58cc09fba0d518254abc0d6150ec262
+size 37801344
diff --git a/backward/L8/100M/sparsity.safetensors b/backward/L8/100M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..dd66a17df660aced8ff6cf40807a0ed445a5cf55
--- /dev/null
+++ b/backward/L8/100M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8b992501a5f990ac63c38f3d1056c26dbd76feb5be2965d9a48981d77940215b
+size 24656
diff --git a/backward/L8/200M/cfg.json b/backward/L8/200M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..7a31d30ec1d468bdacdb4b32d12c742f8842e980
--- /dev/null
+++ b/backward/L8/200M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.8.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 8, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L9", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L8_hook_resid_post_L1_1", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/djw9vsut", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/backward/L8/200M/sae_weights.safetensors b/backward/L8/200M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..70a59f514c3c7147c80324601b8120026ed686fb
--- /dev/null
+++ b/backward/L8/200M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3bfdab5cd4df36ad506e336e3fe8b54514a3870a4ce5c2425994e5bb50207b4d
+size 37801344
diff --git a/backward/L8/200M/sparsity.safetensors b/backward/L8/200M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..87b024d02d2cb4b3a060a2c163c31b6bcb5f81cc
--- /dev/null
+++ b/backward/L8/200M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:4c187b98557ddbc032ad09c447ddb0f77c2fe6d542e843386277bc82cca99118
+size 24656
diff --git a/backward/L8/300M/cfg.json b/backward/L8/300M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..7a31d30ec1d468bdacdb4b32d12c742f8842e980
--- /dev/null
+++ b/backward/L8/300M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.8.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 8, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L9", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L8_hook_resid_post_L1_1", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/djw9vsut", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/backward/L8/300M/sae_weights.safetensors b/backward/L8/300M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..5d73d5e4553c0b4bab764083ea87914da804e4e9
--- /dev/null
+++ b/backward/L8/300M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3ac0d2b9f0886da1b393d1b148c06643d75206bc106c77a5201a98e22e5c2e06
+size 37801344
diff --git a/backward/L8/300M/sparsity.safetensors b/backward/L8/300M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..8b7b66d9441543c2658917b7a1c93d6cb17ade7e
--- /dev/null
+++ b/backward/L8/300M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:057e839a6f93fa3905a99dc4dd65357de53b5ad3884c032e8d3723cd89c47ff8
+size 24656
diff --git a/backward/L8/400M/cfg.json b/backward/L8/400M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..7a31d30ec1d468bdacdb4b32d12c742f8842e980
--- /dev/null
+++ b/backward/L8/400M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.8.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 8, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L9", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L8_hook_resid_post_L1_1", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/djw9vsut", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/backward/L8/400M/sae_weights.safetensors b/backward/L8/400M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..496d6ab0aaae92ec58de37c54dcb4944d10e1a25
--- /dev/null
+++ b/backward/L8/400M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c8d17604d5fb6be2b153dc52f6ad34fac7ab51eced26c5c059d395b88d582365
+size 37801344
diff --git a/backward/L8/400M/sparsity.safetensors b/backward/L8/400M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..a73621cdd853a493fde11ef5106c00752f66c9cb
--- /dev/null
+++ b/backward/L8/400M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b9baca68defa7b370c8f44378abc7cfc1c800cc75ca3649cb30a703346ced688
+size 24656
diff --git a/backward/L8/500M/cfg.json b/backward/L8/500M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..7a31d30ec1d468bdacdb4b32d12c742f8842e980
--- /dev/null
+++ b/backward/L8/500M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.8.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 8, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L9", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L8_hook_resid_post_L1_1", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/djw9vsut", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/backward/L8/500M/sae_weights.safetensors b/backward/L8/500M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..2c03e761ff3144c9360b63a0fc0e8cedae482c41
--- /dev/null
+++ b/backward/L8/500M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:57d935477552535a76aa373274260a92af693dacb7b41e8fd7c7b74b6c69fbd1
+size 37801344
diff --git a/backward/L8/500M/sparsity.safetensors b/backward/L8/500M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..af6c84a6a61e32000e202dc80e5e2e86d62ca995
--- /dev/null
+++ b/backward/L8/500M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c6c2c2577594fdda2908b675580ad41cd6f4672a7dfbcf94f7da6159553bfb1e
+size 24656
diff --git a/backward/L9/100M/cfg.json b/backward/L9/100M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..b135ac4313f8fe9ad24b8838e09d726e6a34354f
--- /dev/null
+++ b/backward/L9/100M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.9.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 9, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L10", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L9_hook_resid_post_L1_1", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/drbi3oow", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/backward/L9/100M/sae_weights.safetensors b/backward/L9/100M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..b5e0d3f1dab17b5ce836bbfce22ff2e1c6308e30
--- /dev/null
+++ b/backward/L9/100M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:240a5ea1acd913b121e20e76ed4713fe80fa907c8d69178ca37ea9a2f9e21329
+size 37801344
diff --git a/backward/L9/100M/sparsity.safetensors b/backward/L9/100M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..310d23b6e43d32a78d9afda69b67f2f9030c29d0
--- /dev/null
+++ b/backward/L9/100M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c8bd076cb5293df80bf5138602ddb0e898ed59500252f5346460a37a0f234813
+size 24656
diff --git a/backward/L9/200M/cfg.json b/backward/L9/200M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..b135ac4313f8fe9ad24b8838e09d726e6a34354f
--- /dev/null
+++ b/backward/L9/200M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.9.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 9, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L10", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L9_hook_resid_post_L1_1", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/drbi3oow", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/backward/L9/200M/sae_weights.safetensors b/backward/L9/200M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..53b75f8bd98927bd0025a0e2d5d54878743f4765
--- /dev/null
+++ b/backward/L9/200M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5a080bb6c343a16bb1a19f7bdb1d8c4e271a29754e6540ba49df08746ed0bb2e
+size 37801344
diff --git a/backward/L9/200M/sparsity.safetensors b/backward/L9/200M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..9d89573855aac289b1f0a918e3d2a0bf6217cb75
--- /dev/null
+++ b/backward/L9/200M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8ec38a833b90d61e9e6aede615c0a34c5b34deb6ba5b896689193e33887788cd
+size 24656
diff --git a/backward/L9/300M/cfg.json b/backward/L9/300M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..b135ac4313f8fe9ad24b8838e09d726e6a34354f
--- /dev/null
+++ b/backward/L9/300M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.9.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 9, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L10", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L9_hook_resid_post_L1_1", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/drbi3oow", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/backward/L9/300M/sae_weights.safetensors b/backward/L9/300M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..8794c543ee0de30271dcf42a5e63ae02649bda92
--- /dev/null
+++ b/backward/L9/300M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d663f5d6adbce65785a58b21600505269afe9aacd7052753f8f5d26af74b07e3
+size 37801344
diff --git a/backward/L9/300M/sparsity.safetensors b/backward/L9/300M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..8ccaf2daa62f866131cd25e5f55736cc17d24ee2
--- /dev/null
+++ b/backward/L9/300M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e9924396a639a0b40eb0cd5821c14f1403e3902d93187ed14c090666a41cb9ec
+size 24656
diff --git a/backward/L9/400M/cfg.json b/backward/L9/400M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..b135ac4313f8fe9ad24b8838e09d726e6a34354f
--- /dev/null
+++ b/backward/L9/400M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.9.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 9, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L10", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L9_hook_resid_post_L1_1", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/drbi3oow", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/backward/L9/400M/sae_weights.safetensors b/backward/L9/400M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..a03f9b3e1799e7bf46ae336a0f24a4c5c42dc54e
--- /dev/null
+++ b/backward/L9/400M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1927ac7c1727f991bf36c5754bee0b69db35e5c59b44fa8aef749a66d7c29612
+size 37801344
diff --git a/backward/L9/400M/sparsity.safetensors b/backward/L9/400M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..1a4f3180097db72e5ec1f876b246c8dcc733bf08
--- /dev/null
+++ b/backward/L9/400M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6ec24b0de2e36541fb6d0918004cb1112990e3fe0968433be5ac9376db70b244
+size 24656
diff --git a/backward/L9/500M/cfg.json b/backward/L9/500M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..b135ac4313f8fe9ad24b8838e09d726e6a34354f
--- /dev/null
+++ b/backward/L9/500M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.9.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 9, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L10", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L9_hook_resid_post_L1_1", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/drbi3oow", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/backward/L9/500M/sae_weights.safetensors b/backward/L9/500M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..a2bc01b0ef979887b0ba5213e88b82f5df7cd724
--- /dev/null
+++ b/backward/L9/500M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:834432085e0d15a45c3c2311f52c3546a98eb522a7d52c4eeb3d9086a807ea7e
+size 37801344
diff --git a/backward/L9/500M/sparsity.safetensors b/backward/L9/500M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..7e21268f3cba8aff303dba88822d3c3fd6149425
--- /dev/null
+++ b/backward/L9/500M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8740f1fae01c9d33e78e74f8ca20b6ef4b080db496e476ec16a312801a962f0b
+size 24656
diff --git a/forward/L1/100M/cfg.json b/forward/L1/100M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..925a86ee8166d079b62c5f8e5cb44c8136fe1562
--- /dev/null
+++ b/forward/L1/100M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.1.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 1, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L0", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L1_hook_resid_post_L1_1_FW", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/i0l4e4xt", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/forward/L1/100M/sae_weights.safetensors b/forward/L1/100M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..86b0b37cea0eb5ec3292aa91fcd2c3e5c237c075
--- /dev/null
+++ b/forward/L1/100M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ba032b6b1e0c4f38e41078df6d84a715f3c6554c932e70742b623b77bd7883b8
+size 37801344
diff --git a/forward/L1/100M/sparsity.safetensors b/forward/L1/100M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..bafa4e67ff39daf8486258e6c9d24a42d4831d98
--- /dev/null
+++ b/forward/L1/100M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:495f0db107e09e27fd8d9a98f91373c31887e0f4e51fa44a8beae844556e67c7
+size 24656
diff --git a/forward/L1/200M/cfg.json b/forward/L1/200M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..925a86ee8166d079b62c5f8e5cb44c8136fe1562
--- /dev/null
+++ b/forward/L1/200M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.1.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 1, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L0", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L1_hook_resid_post_L1_1_FW", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/i0l4e4xt", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/forward/L1/200M/sae_weights.safetensors b/forward/L1/200M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..82d07c20c4cf2515c06f6a1ca3bae5de2ddae564
--- /dev/null
+++ b/forward/L1/200M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f928f0bddd1547e3fd4185e1d4cb65d976d8b3d80e8e86b77ab63f443f7d60eb
+size 37801344
diff --git a/forward/L1/200M/sparsity.safetensors b/forward/L1/200M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..5bce7160c9dcb8b507d7f11745c3ddc0d4230787
--- /dev/null
+++ b/forward/L1/200M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5cab824cc3e6c1f3da1778865a5f235ce554051a20d0097781e8211c51a6c9b0
+size 24656
diff --git a/forward/L1/300M/cfg.json b/forward/L1/300M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..925a86ee8166d079b62c5f8e5cb44c8136fe1562
--- /dev/null
+++ b/forward/L1/300M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.1.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 1, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L0", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L1_hook_resid_post_L1_1_FW", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/i0l4e4xt", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/forward/L1/300M/sae_weights.safetensors b/forward/L1/300M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..a28fe02073cf5cd691f999e508dfc6b71b553786
--- /dev/null
+++ b/forward/L1/300M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:cdc656580f65f5eca1ac78bbf3b3b0c17c4fcc94e6c47e7aa3862892788bb552
+size 37801344
diff --git a/forward/L1/300M/sparsity.safetensors b/forward/L1/300M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..e931feb95561b951d8d885293a51db27ecd10dfc
--- /dev/null
+++ b/forward/L1/300M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7bde7d70af092d02397a4d9140395508ea072f5019058eddc2e17a9905fb48c9
+size 24656
diff --git a/forward/L1/400M/cfg.json b/forward/L1/400M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..925a86ee8166d079b62c5f8e5cb44c8136fe1562
--- /dev/null
+++ b/forward/L1/400M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.1.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 1, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L0", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L1_hook_resid_post_L1_1_FW", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/i0l4e4xt", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/forward/L1/400M/sae_weights.safetensors b/forward/L1/400M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..5ad72b3b13eff3c3477f6b80c73f5435f46d7706
--- /dev/null
+++ b/forward/L1/400M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b9ce976a959643ea24942634413949867fddbc4daeca530d064b26dd63511af2
+size 37801344
diff --git a/forward/L1/400M/sparsity.safetensors b/forward/L1/400M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..dbeae27e69990423aef9c929d9f198dabfe03c78
--- /dev/null
+++ b/forward/L1/400M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2d064cdffa9db5835fb1ce1c2497590f657e632c2756630caf85331d9a225e5c
+size 24656
diff --git a/forward/L1/500M/cfg.json b/forward/L1/500M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..925a86ee8166d079b62c5f8e5cb44c8136fe1562
--- /dev/null
+++ b/forward/L1/500M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.1.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 1, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L0", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L1_hook_resid_post_L1_1_FW", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/i0l4e4xt", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/forward/L1/500M/sae_weights.safetensors b/forward/L1/500M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..f9f10ebbf3d3bbb63f33cbd1f89b1c35add1803a
--- /dev/null
+++ b/forward/L1/500M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:85c13de52358ba686531ea108bca4154b8b30fb6f05cb59942aca5f98984944e
+size 37801344
diff --git a/forward/L1/500M/sparsity.safetensors b/forward/L1/500M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..e651134a1d572ee8a1fa55dd0769a7f7c21d5d7f
--- /dev/null
+++ b/forward/L1/500M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9bb17c6eb7754087668acfd2858a3da4ab7817df9e1c4c33c38bc34678f74289
+size 24656
diff --git a/forward/L10/100M/cfg.json b/forward/L10/100M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..8cebb6590074843e80dbf7e6efa86d1dd86efbf6
--- /dev/null
+++ b/forward/L10/100M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.10.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 10, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L9", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L10_hook_resid_post_L1_1_FW", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/rkol4qfk", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/forward/L10/100M/sae_weights.safetensors b/forward/L10/100M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..659d459d6117fa43434a8b71e5b36e55fbc7dfe7
--- /dev/null
+++ b/forward/L10/100M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:19174cb1e7b07c06c5b79ae4ff449e44b8d3a038d3cf155a011833420117ec19
+size 37801344
diff --git a/forward/L10/100M/sparsity.safetensors b/forward/L10/100M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..4e75823b3135968f30ca483853af92b979455b17
--- /dev/null
+++ b/forward/L10/100M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9517c5409c5df880e4f89381b16af45bddd7b6274a05082d553a3b7640a36728
+size 24656
diff --git a/forward/L10/200M/cfg.json b/forward/L10/200M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..8cebb6590074843e80dbf7e6efa86d1dd86efbf6
--- /dev/null
+++ b/forward/L10/200M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.10.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 10, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L9", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L10_hook_resid_post_L1_1_FW", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/rkol4qfk", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/forward/L10/200M/sae_weights.safetensors b/forward/L10/200M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..47284f228c0a05ca1be77a13d64d5e9a3319e98e
--- /dev/null
+++ b/forward/L10/200M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d18690a479b9be07b5af57e5d7235b616bca244b8ffe9aa01183d0e29c90a130
+size 37801344
diff --git a/forward/L10/200M/sparsity.safetensors b/forward/L10/200M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..62649cf0a7402036060fb19986bf43894f2d2b14
--- /dev/null
+++ b/forward/L10/200M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2fedbb499448f0fc4a772dfff38df8ab0fda3e2ed25eb7505a44084247a13e47
+size 24656
diff --git a/forward/L10/300M/cfg.json b/forward/L10/300M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..8cebb6590074843e80dbf7e6efa86d1dd86efbf6
--- /dev/null
+++ b/forward/L10/300M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.10.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 10, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L9", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L10_hook_resid_post_L1_1_FW", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/rkol4qfk", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/forward/L10/300M/sae_weights.safetensors b/forward/L10/300M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..f121ee763c1c5d2765826301c076a932c4aa673f
--- /dev/null
+++ b/forward/L10/300M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:00b83eec62725089b1f2efeb73f454c638771e970bb0c634bce394863ae0ee93
+size 37801344
diff --git a/forward/L10/300M/sparsity.safetensors b/forward/L10/300M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..87d4b0d2572a2ad331e17d1b4abf3f3146eaa3b9
--- /dev/null
+++ b/forward/L10/300M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ef8af6a7a080b85f742b0b3e7c1048457d5ed26ec1873e61f37c263c1c4077b5
+size 24656
diff --git a/forward/L10/400M/cfg.json b/forward/L10/400M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..8cebb6590074843e80dbf7e6efa86d1dd86efbf6
--- /dev/null
+++ b/forward/L10/400M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.10.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 10, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L9", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L10_hook_resid_post_L1_1_FW", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/rkol4qfk", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/forward/L10/400M/sae_weights.safetensors b/forward/L10/400M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..d851e1ff4a56678bedc6020050f6b6b8e2395f37
--- /dev/null
+++ b/forward/L10/400M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e7c4128c589d6bf1ab0fe3973ab5c4e2ce722fd20f024b2e8bd516746aa62d8e
+size 37801344
diff --git a/forward/L10/400M/sparsity.safetensors b/forward/L10/400M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..8bd3ac9ac186e8806d9dd1717ce99ff9438a534f
--- /dev/null
+++ b/forward/L10/400M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:498f19846654ed26528de8087722cbad0457c6935d84bab16222ada3e60c9513
+size 24656
diff --git a/forward/L10/500M/cfg.json b/forward/L10/500M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..8cebb6590074843e80dbf7e6efa86d1dd86efbf6
--- /dev/null
+++ b/forward/L10/500M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.10.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 10, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L9", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L10_hook_resid_post_L1_1_FW", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/rkol4qfk", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/forward/L10/500M/sae_weights.safetensors b/forward/L10/500M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..1a94c3bb74620efe9dd33a761d8645a31a2111c6
--- /dev/null
+++ b/forward/L10/500M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bab11afb173b989dd12089c9135d25ee5b7b4c6203c99e43845a20d121bd7ea3
+size 37801344
diff --git a/forward/L10/500M/sparsity.safetensors b/forward/L10/500M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..1fce58f9d529891464b2cee6b9cd508ec2f0e512
--- /dev/null
+++ b/forward/L10/500M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a10db6a532c3479637ce72eaf08029de0b8ac6133ac7d1f9be8bd39a1e478d83
+size 24656
diff --git a/forward/L11/100M/cfg.json b/forward/L11/100M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..c1ce1605d31c08f0022d421b088629be6af19d0a
--- /dev/null
+++ b/forward/L11/100M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.11.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 11, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L10", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L11_hook_resid_post_L1_1_FW", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/b6n3tsnf", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/forward/L11/100M/sae_weights.safetensors b/forward/L11/100M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..c2ec6cc3cce14ac9ab0e73166eeb18c7264f59c2
--- /dev/null
+++ b/forward/L11/100M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6adea5bd9a8645fe0cd881e55de4a9ce6ca8b2d09829e67f6a1e952ba3d36b83
+size 37801344
diff --git a/forward/L11/100M/sparsity.safetensors b/forward/L11/100M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..1efd6fc071c4e74f0001384138d722977704f127
--- /dev/null
+++ b/forward/L11/100M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:abe802a8188fe084c69aea93c99812d0d6b7a4406e4cf5f860bd484b5a073449
+size 24656
diff --git a/forward/L11/200M/cfg.json b/forward/L11/200M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..c1ce1605d31c08f0022d421b088629be6af19d0a
--- /dev/null
+++ b/forward/L11/200M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.11.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 11, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L10", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L11_hook_resid_post_L1_1_FW", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/b6n3tsnf", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/forward/L11/200M/sae_weights.safetensors b/forward/L11/200M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..7813765bf06d85216a1459c5ebadeb55459a27c7
--- /dev/null
+++ b/forward/L11/200M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0eea23cd1997e30d5c8b86fddbf7405cc54d5a5fc3eebc29236fade4c3c725bb
+size 37801344
diff --git a/forward/L11/200M/sparsity.safetensors b/forward/L11/200M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..8325fcda0cf3eab6d8bf3e4e48fc02ad746e6f90
--- /dev/null
+++ b/forward/L11/200M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:74ae2659c67e2408ccdc44b30da31289656fde82f8e9329d0d0d879ed9495622
+size 24656
diff --git a/forward/L11/300M/cfg.json b/forward/L11/300M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..c1ce1605d31c08f0022d421b088629be6af19d0a
--- /dev/null
+++ b/forward/L11/300M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.11.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 11, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L10", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L11_hook_resid_post_L1_1_FW", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/b6n3tsnf", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/forward/L11/300M/sae_weights.safetensors b/forward/L11/300M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..f0196fe12403025ddde2f73947c8384930d4c23b
--- /dev/null
+++ b/forward/L11/300M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ddf2aaea67f7a301ade94f89fefd0ddb6f77967cef932fc01c4f9127944a05b1
+size 37801344
diff --git a/forward/L11/300M/sparsity.safetensors b/forward/L11/300M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..758b29d265bee0716131e00dbea7f72b77db8502
--- /dev/null
+++ b/forward/L11/300M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:054bd93c3f13b2cfaca585ec810d85f983057fa9567a94dfa70e7ba19d0676ea
+size 24656
diff --git a/forward/L11/400M/cfg.json b/forward/L11/400M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..c1ce1605d31c08f0022d421b088629be6af19d0a
--- /dev/null
+++ b/forward/L11/400M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.11.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 11, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L10", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L11_hook_resid_post_L1_1_FW", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/b6n3tsnf", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/forward/L11/400M/sae_weights.safetensors b/forward/L11/400M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..a97464ab361ad9e37ebb7aaf342452d7310dd5d3
--- /dev/null
+++ b/forward/L11/400M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:aca7570aec8770477d6b8a086778b8abea6f765046a106d029aab7891a8888fa
+size 37801344
diff --git a/forward/L11/400M/sparsity.safetensors b/forward/L11/400M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..c82b647b3ac6f68685677309e3d97546075357f9
--- /dev/null
+++ b/forward/L11/400M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:42d7f8dac03916b330f2a5521296f4f744589d82a19a2ce90ccbae8ae07e9cd9
+size 24656
diff --git a/forward/L11/500M/cfg.json b/forward/L11/500M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..c1ce1605d31c08f0022d421b088629be6af19d0a
--- /dev/null
+++ b/forward/L11/500M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.11.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 11, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L10", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L11_hook_resid_post_L1_1_FW", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/b6n3tsnf", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/forward/L11/500M/sae_weights.safetensors b/forward/L11/500M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..a057a4e51d7432a1509c69676716c7decfe16674
--- /dev/null
+++ b/forward/L11/500M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:be193dee14169955d78669513b5a2370a7e27bedda1fdcd18fb57cfcad09450b
+size 37801344
diff --git a/forward/L11/500M/sparsity.safetensors b/forward/L11/500M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..619bd9ce7cb30207c1cb5c82c8dbaced82b6b91d
--- /dev/null
+++ b/forward/L11/500M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:11cf36b0a5546fa71450bb30ee9fa0007b84b9a30b4f813c700a81abce61a3b0
+size 24656
diff --git a/forward/L2/100M/cfg.json b/forward/L2/100M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..9544ecd9aef8c07108decf0d5698c645b0029b9e
--- /dev/null
+++ b/forward/L2/100M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.2.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 2, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L1", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L2_hook_resid_post_L1_1_FW", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/ika13rog", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/forward/L2/100M/sae_weights.safetensors b/forward/L2/100M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..b47ecc17b94505993d8223f4c4fef67838d4302d
--- /dev/null
+++ b/forward/L2/100M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:44a90b5d71c664cb4500a686118528ff47f3267dbfc79b1b5552500334a6ea95
+size 37801344
diff --git a/forward/L2/100M/sparsity.safetensors b/forward/L2/100M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..96e9da74be5cd606307f6cecc5127ec41fb592d2
--- /dev/null
+++ b/forward/L2/100M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1254655de7a8f9966c9a690bac547e7892d8f77be4cb07771ae6f93dceb0d367
+size 24656
diff --git a/forward/L2/200M/cfg.json b/forward/L2/200M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..9544ecd9aef8c07108decf0d5698c645b0029b9e
--- /dev/null
+++ b/forward/L2/200M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.2.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 2, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L1", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L2_hook_resid_post_L1_1_FW", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/ika13rog", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/forward/L2/200M/sae_weights.safetensors b/forward/L2/200M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..1e6ab05ccb252beb99aa33fc13bd5b8e7b139ab9
--- /dev/null
+++ b/forward/L2/200M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c9807fb7a408869728cdd6d061aa15c81b91b11aae143b1ac558e585af1576d7
+size 37801344
diff --git a/forward/L2/200M/sparsity.safetensors b/forward/L2/200M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..a0ae98aac112b260d9bbdbcecfd13de35228fec7
--- /dev/null
+++ b/forward/L2/200M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:83603ff9e65607ac6f846948ffc9fed57162b8e424c180a2c5a302a898632a0b
+size 24656
diff --git a/forward/L2/300M/cfg.json b/forward/L2/300M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..9544ecd9aef8c07108decf0d5698c645b0029b9e
--- /dev/null
+++ b/forward/L2/300M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.2.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 2, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L1", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L2_hook_resid_post_L1_1_FW", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/ika13rog", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/forward/L2/300M/sae_weights.safetensors b/forward/L2/300M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..2a019b7adcaea67b9d02fbc5eafbfc95c1e5e491
--- /dev/null
+++ b/forward/L2/300M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:eaad3689815f0b9f27c9491b66874dbd76dd2d11aee5efa16706dae6db31eaed
+size 37801344
diff --git a/forward/L2/300M/sparsity.safetensors b/forward/L2/300M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..54d97e04cfa8d89dec3c480d46c4f40fbd27046c
--- /dev/null
+++ b/forward/L2/300M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5904793dfbb3823ee6d4af6dc3806db128ba5a947bacfea1217c214cfdb1bb84
+size 24656
diff --git a/forward/L2/400M/cfg.json b/forward/L2/400M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..9544ecd9aef8c07108decf0d5698c645b0029b9e
--- /dev/null
+++ b/forward/L2/400M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.2.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 2, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L1", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L2_hook_resid_post_L1_1_FW", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/ika13rog", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/forward/L2/400M/sae_weights.safetensors b/forward/L2/400M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..5c9dbdfe633a2f9c4132d171157986491408e7f9
--- /dev/null
+++ b/forward/L2/400M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:053b977f4591e1d64dc9ba47fcc82fa3c4c701d8feef6b9708e01bcaf9b38e1e
+size 37801344
diff --git a/forward/L2/400M/sparsity.safetensors b/forward/L2/400M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..fe655e5c976e9ca2857c0ace2ee6477c2253a49d
--- /dev/null
+++ b/forward/L2/400M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c429b755e4ac87cb267c6ad47a226ebcddbe8bf640eeae47dfc155bfe0962c24
+size 24656
diff --git a/forward/L2/500M/cfg.json b/forward/L2/500M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..9544ecd9aef8c07108decf0d5698c645b0029b9e
--- /dev/null
+++ b/forward/L2/500M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.2.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 2, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L1", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L2_hook_resid_post_L1_1_FW", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/ika13rog", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/forward/L2/500M/sae_weights.safetensors b/forward/L2/500M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..f86e9413a67507b2aee3486e8266afa8aab757b6
--- /dev/null
+++ b/forward/L2/500M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c58746c7b76583c1a83801822ecb7c1a0ce4b43fcae64858acab4ee638c093de
+size 37801344
diff --git a/forward/L2/500M/sparsity.safetensors b/forward/L2/500M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..cd0a884f2d17144fb06b0f254a7fade6e7a13171
--- /dev/null
+++ b/forward/L2/500M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c715492446a7d081b53f563d7e90dcfd655cc3277bc8c1bcd2934ce4e524cac0
+size 24656
diff --git a/forward/L3/100M/cfg.json b/forward/L3/100M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..a1c1e17edbb2b1c2086a55ec41937be26284701c
--- /dev/null
+++ b/forward/L3/100M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.3.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 3, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L2", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L3_hook_resid_post_L1_1_FW", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/z9wc7uv1", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/forward/L3/100M/sae_weights.safetensors b/forward/L3/100M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..6a49daaadde8a6995f8b85123cb42cb0913c2a74
--- /dev/null
+++ b/forward/L3/100M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6bb03d56371a7420294965a8eb5d208b9c4f95d137bc6bbdabfdc16327179d75
+size 37801344
diff --git a/forward/L3/100M/sparsity.safetensors b/forward/L3/100M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..eec05b799b0b2a7d01715b960e155383970195cd
--- /dev/null
+++ b/forward/L3/100M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b2a36bd39d99f1a6ed748a4c721733d8250f221c6e29551412ed745d159c2266
+size 24656
diff --git a/forward/L3/200M/cfg.json b/forward/L3/200M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..a1c1e17edbb2b1c2086a55ec41937be26284701c
--- /dev/null
+++ b/forward/L3/200M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.3.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 3, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L2", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L3_hook_resid_post_L1_1_FW", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/z9wc7uv1", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/forward/L3/200M/sae_weights.safetensors b/forward/L3/200M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..9e607a4cd0ae1c45b4791c21df5209cf791badb4
--- /dev/null
+++ b/forward/L3/200M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8c183e608c3988842f883ee68fe656f4a933fc1cad01feec7193020fe798b019
+size 37801344
diff --git a/forward/L3/200M/sparsity.safetensors b/forward/L3/200M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..51fc2ef00a3809831e95fba0b4b8ef661a40e2a7
--- /dev/null
+++ b/forward/L3/200M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ef982d2fbb09c41843f6f3afd7c7368336356996e6dbae8857a3989e4a7f1266
+size 24656
diff --git a/forward/L3/300M/cfg.json b/forward/L3/300M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..a1c1e17edbb2b1c2086a55ec41937be26284701c
--- /dev/null
+++ b/forward/L3/300M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.3.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 3, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L2", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L3_hook_resid_post_L1_1_FW", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/z9wc7uv1", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/forward/L3/300M/sae_weights.safetensors b/forward/L3/300M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..d92e546be0dbd72a2a1848b6ed732a30647d06bf
--- /dev/null
+++ b/forward/L3/300M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3cf94d831865b2da7c720e4b2452d2ee4bb15e8fa1477c609ca11b1c327db967
+size 37801344
diff --git a/forward/L3/300M/sparsity.safetensors b/forward/L3/300M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..603cc92e5d7f4952c859fcf73370d1d349d2044b
--- /dev/null
+++ b/forward/L3/300M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a9c4efcc4a5f6be7a35e6f8de0406f29a47cef842f6f3c1d781e046103b9335e
+size 24656
diff --git a/forward/L3/400M/cfg.json b/forward/L3/400M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..a1c1e17edbb2b1c2086a55ec41937be26284701c
--- /dev/null
+++ b/forward/L3/400M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.3.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 3, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L2", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L3_hook_resid_post_L1_1_FW", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/z9wc7uv1", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/forward/L3/400M/sae_weights.safetensors b/forward/L3/400M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..f896f66ac77da58119f99798a523e88bb6811cc2
--- /dev/null
+++ b/forward/L3/400M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:43ed4b1e02f3111a99f577ae41a0d530ce474eaf2fea1da0254800e1c138dd82
+size 37801344
diff --git a/forward/L3/400M/sparsity.safetensors b/forward/L3/400M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..711151086d6b7e61f6507dae26e30db1b8672d20
--- /dev/null
+++ b/forward/L3/400M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5b5cc7667b574158a532113d20af1a5b0427b36d836cece85fcb0eb52faa048c
+size 24656
diff --git a/forward/L3/500M/cfg.json b/forward/L3/500M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..a1c1e17edbb2b1c2086a55ec41937be26284701c
--- /dev/null
+++ b/forward/L3/500M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.3.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 3, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L2", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L3_hook_resid_post_L1_1_FW", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/z9wc7uv1", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/forward/L3/500M/sae_weights.safetensors b/forward/L3/500M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..8c2f16a917593bca3a48948cdb4356266176f5d4
--- /dev/null
+++ b/forward/L3/500M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d85e6139535feab8a3ed66c48e8c1500939923f843547ef51bda6e4b14f70cd4
+size 37801344
diff --git a/forward/L3/500M/sparsity.safetensors b/forward/L3/500M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..1e4ed9a2253d8d5b5283b05696fd8f822668861e
--- /dev/null
+++ b/forward/L3/500M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:704f7ee206c2dfdd7e3a326514d5ce26a679f5097e77e3b3fb6bff5663543906
+size 24656
diff --git a/forward/L4/100M/cfg.json b/forward/L4/100M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..d146d329ee66ac141372f1e8bb3494776498f772
--- /dev/null
+++ b/forward/L4/100M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.4.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 4, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L3", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L4_hook_resid_post_L1_1_FW", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/6s2vc5va", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/forward/L4/100M/sae_weights.safetensors b/forward/L4/100M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..7fe1aeed2e272e7b70e2d03df75adcd77438129f
--- /dev/null
+++ b/forward/L4/100M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b1c6fe2415252cf190af2cfa66def158aa016efe048d9118a6a9e01cce07bd19
+size 37801344
diff --git a/forward/L4/100M/sparsity.safetensors b/forward/L4/100M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..d5a7ee64cd8c8fad178bcde36f3cb3c6f740aef2
--- /dev/null
+++ b/forward/L4/100M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:154ec55449800692839c1e332a94beaf8c9260b5a0e6d7757ead91d159431116
+size 24656
diff --git a/forward/L4/200M/cfg.json b/forward/L4/200M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..d146d329ee66ac141372f1e8bb3494776498f772
--- /dev/null
+++ b/forward/L4/200M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.4.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 4, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L3", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L4_hook_resid_post_L1_1_FW", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/6s2vc5va", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/forward/L4/200M/sae_weights.safetensors b/forward/L4/200M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..a26d8312663b8a47c484392e1f73a3246954164d
--- /dev/null
+++ b/forward/L4/200M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:479e8d53a4de3697a9e1aa23ac60d8d5e9fdd2352c8ba9bdc39844e3c735a75f
+size 37801344
diff --git a/forward/L4/200M/sparsity.safetensors b/forward/L4/200M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..fa70df148c99424a8d85b464a2adfdbc3e800290
--- /dev/null
+++ b/forward/L4/200M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f77c97270ae3582c849a2c209918bcb47d3d3f05d07d4433ec0a04c8278815d1
+size 24656
diff --git a/forward/L4/300M/cfg.json b/forward/L4/300M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..d146d329ee66ac141372f1e8bb3494776498f772
--- /dev/null
+++ b/forward/L4/300M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.4.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 4, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L3", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L4_hook_resid_post_L1_1_FW", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/6s2vc5va", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/forward/L4/300M/sae_weights.safetensors b/forward/L4/300M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..1142346c6f318efdcebf2da9207ce0475c6f8647
--- /dev/null
+++ b/forward/L4/300M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0ffa323947044fbc7d9eb21a8b4e7f6213db0e07f46fd0e1c85aa52405ad03b1
+size 37801344
diff --git a/forward/L4/300M/sparsity.safetensors b/forward/L4/300M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..544c2a1149b4e317579543456ebd373733998df8
--- /dev/null
+++ b/forward/L4/300M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6f452a250d1c1a864cee7ad8289c33de4b95626ae6d416c9bf32c631f0db37ba
+size 24656
diff --git a/forward/L4/400M/cfg.json b/forward/L4/400M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..d146d329ee66ac141372f1e8bb3494776498f772
--- /dev/null
+++ b/forward/L4/400M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.4.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 4, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L3", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L4_hook_resid_post_L1_1_FW", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/6s2vc5va", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/forward/L4/400M/sae_weights.safetensors b/forward/L4/400M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..eeb3c1cc5796bbfaf5af590604d704e3ef64def8
--- /dev/null
+++ b/forward/L4/400M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a2b4bc50fb1aff654e724e2a9f0bfa38009837ebe9a6e09ae739475fc3702431
+size 37801344
diff --git a/forward/L4/400M/sparsity.safetensors b/forward/L4/400M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..ef99d0673152f624e7229fa74cc2f64adde2a820
--- /dev/null
+++ b/forward/L4/400M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:04e1d336928e92cc0c050e1bc74928cdcecf13704c6585320043e7e505cd9512
+size 24656
diff --git a/forward/L4/500M/cfg.json b/forward/L4/500M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..d146d329ee66ac141372f1e8bb3494776498f772
--- /dev/null
+++ b/forward/L4/500M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.4.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 4, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L3", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L4_hook_resid_post_L1_1_FW", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/6s2vc5va", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/forward/L4/500M/sae_weights.safetensors b/forward/L4/500M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..0881400838527041a7b259ce469e587c2d00ace3
--- /dev/null
+++ b/forward/L4/500M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0a34ff9264370aaba2935d6d2943e4e94e460babb7385b60d9765ad6e5a13545
+size 37801344
diff --git a/forward/L4/500M/sparsity.safetensors b/forward/L4/500M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..907ee9cc95568fba4f20a5e2d25f8dfc71a606a5
--- /dev/null
+++ b/forward/L4/500M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:90daeec0a1a7e4dca8267f35171f84c61d99a08d356411ee37248e75f1f4eade
+size 24656
diff --git a/forward/L5/100M/cfg.json b/forward/L5/100M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..ba3c05e914542f076d07678e08a7faf551e0985a
--- /dev/null
+++ b/forward/L5/100M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.5.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 5, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L4", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L5_hook_resid_post_L1_1_FW", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/rxc0vgj7", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/forward/L5/100M/sae_weights.safetensors b/forward/L5/100M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..0bda925ccd964b15f9926f4a4741f2156ca3f9cd
--- /dev/null
+++ b/forward/L5/100M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a54c6390949262bb0f86a7344111475e213b3919a756f3dd4d1ab4659d5f5b2a
+size 37801344
diff --git a/forward/L5/100M/sparsity.safetensors b/forward/L5/100M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..c781eb7900a97d208d1761dfd38d914a9360b95d
--- /dev/null
+++ b/forward/L5/100M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:926b113f23acc9ee9650a2aee80970cab9ab4fa65dd58dd8764581d7653f2491
+size 24656
diff --git a/forward/L5/200M/cfg.json b/forward/L5/200M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..ba3c05e914542f076d07678e08a7faf551e0985a
--- /dev/null
+++ b/forward/L5/200M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.5.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 5, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L4", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L5_hook_resid_post_L1_1_FW", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/rxc0vgj7", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/forward/L5/200M/sae_weights.safetensors b/forward/L5/200M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..b1bdbd53aafb8f19eaa0649e1e53ed30743d1904
--- /dev/null
+++ b/forward/L5/200M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d012f6b50a86d8a8a9e97c28de1ef391b30dd3873ef03fda9bf229dfb76a725e
+size 37801344
diff --git a/forward/L5/200M/sparsity.safetensors b/forward/L5/200M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..a13cd35f7c81e2de5ff514d75dc83f07a9e1e87f
--- /dev/null
+++ b/forward/L5/200M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ba9300ed77b2e1d6d21ee9fe7ec93563e47b91941904c4513a3268fe60f134e9
+size 24656
diff --git a/forward/L5/300M/cfg.json b/forward/L5/300M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..ba3c05e914542f076d07678e08a7faf551e0985a
--- /dev/null
+++ b/forward/L5/300M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.5.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 5, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L4", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L5_hook_resid_post_L1_1_FW", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/rxc0vgj7", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/forward/L5/300M/sae_weights.safetensors b/forward/L5/300M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..df9d468faa8784323df716f9705338c1bcd26bb6
--- /dev/null
+++ b/forward/L5/300M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c3b382ee76534cbf9d4b23ed10007edb5cd78522e001de6ed6da8ea8bf475311
+size 37801344
diff --git a/forward/L5/300M/sparsity.safetensors b/forward/L5/300M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..abe963e57a211a30efe829d1d5451d62ef5bec3a
--- /dev/null
+++ b/forward/L5/300M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a98ece20390ab2b3136573954a8f870eb71fc919fbf8e6f0d1227710cffb9fe3
+size 24656
diff --git a/forward/L5/400M/cfg.json b/forward/L5/400M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..ba3c05e914542f076d07678e08a7faf551e0985a
--- /dev/null
+++ b/forward/L5/400M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.5.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 5, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L4", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L5_hook_resid_post_L1_1_FW", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/rxc0vgj7", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/forward/L5/400M/sae_weights.safetensors b/forward/L5/400M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..714003afae0f02b85a32d26f8a77a21242a4a767
--- /dev/null
+++ b/forward/L5/400M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6e2cd17d90dc421119812f58011426bbaee73dff46c6eaf63801bd63ad384d25
+size 37801344
diff --git a/forward/L5/400M/sparsity.safetensors b/forward/L5/400M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..a4994559cf5e7c46d2056c7366f2fbe40bb26633
--- /dev/null
+++ b/forward/L5/400M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6e284e615f69b624fcadf499e7ecb51f1972b8042cb1f093c478f9f88c080fd1
+size 24656
diff --git a/forward/L5/500M/cfg.json b/forward/L5/500M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..ba3c05e914542f076d07678e08a7faf551e0985a
--- /dev/null
+++ b/forward/L5/500M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.5.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 5, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L4", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L5_hook_resid_post_L1_1_FW", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/rxc0vgj7", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/forward/L5/500M/sae_weights.safetensors b/forward/L5/500M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..43c6c62a3ed96bd3941d4516c0b0a0f6168f3c92
--- /dev/null
+++ b/forward/L5/500M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:5b68ada82ed460ecf17e41b0db14f93f67d5a59d1051df7a099ef4fe4148307f
+size 37801344
diff --git a/forward/L5/500M/sparsity.safetensors b/forward/L5/500M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..f170e32835f7a6445999083176ef5c20913b19a7
--- /dev/null
+++ b/forward/L5/500M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:20d957f78a14f19c41a041491e864be00499ce3a1cd32391ad0ee5486350cfe0
+size 24656
diff --git a/forward/L6/100M/cfg.json b/forward/L6/100M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..16d9484eac4031b59a29953628b897a659a45bae
--- /dev/null
+++ b/forward/L6/100M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.6.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 6, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L5", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L6_hook_resid_post_L1_1_FW", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/n8vubr98", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/forward/L6/100M/sae_weights.safetensors b/forward/L6/100M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..e0069fdaa6918213d4ec34142301c4fb9d449696
--- /dev/null
+++ b/forward/L6/100M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:531bc922ce2d6a430e54f1de91f4c08ee232ce713f03a51e2b9b9333c883228e
+size 37801344
diff --git a/forward/L6/100M/sparsity.safetensors b/forward/L6/100M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..60275b6bf368ed466e1c9e5e75708ba45c94e97f
--- /dev/null
+++ b/forward/L6/100M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:3c7a54623f7cc683ec604156a07f9f6ea85b886842d0ff15d219bb007abbd23d
+size 24656
diff --git a/forward/L6/200M/cfg.json b/forward/L6/200M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..16d9484eac4031b59a29953628b897a659a45bae
--- /dev/null
+++ b/forward/L6/200M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.6.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 6, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L5", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L6_hook_resid_post_L1_1_FW", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/n8vubr98", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/forward/L6/200M/sae_weights.safetensors b/forward/L6/200M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..40233487d2c42dd23a18b6b1d6ca36196b382b7a
--- /dev/null
+++ b/forward/L6/200M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:aa6e3a1e8f1d9911015bd68a16f9ce95045617d296b69c9ded88d6a49380db71
+size 37801344
diff --git a/forward/L6/200M/sparsity.safetensors b/forward/L6/200M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..56346f06759a166c11ce150c54e8d2db8de66c7b
--- /dev/null
+++ b/forward/L6/200M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d2c4ea7ce72a5af82a8bedcba118b97498f61598e213092a92f1cc69853b956b
+size 24656
diff --git a/forward/L6/300M/cfg.json b/forward/L6/300M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..16d9484eac4031b59a29953628b897a659a45bae
--- /dev/null
+++ b/forward/L6/300M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.6.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 6, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L5", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L6_hook_resid_post_L1_1_FW", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/n8vubr98", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/forward/L6/300M/sae_weights.safetensors b/forward/L6/300M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..ddd5a3dd919bae21e5fd008a33e634b0e365d127
--- /dev/null
+++ b/forward/L6/300M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8261ccb99e6ac4de355601118f13a04c1d5c201404e6b0b5f3e7a545a3e58cf1
+size 37801344
diff --git a/forward/L6/300M/sparsity.safetensors b/forward/L6/300M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..10b90b7f046a209c1cd9e540a9338a350df1c4b0
--- /dev/null
+++ b/forward/L6/300M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:48055c481013717d99177be986c3b0d259a82c65f8ef5624260b7e0bcbf6dd6d
+size 24656
diff --git a/forward/L6/400M/cfg.json b/forward/L6/400M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..16d9484eac4031b59a29953628b897a659a45bae
--- /dev/null
+++ b/forward/L6/400M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.6.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 6, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L5", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L6_hook_resid_post_L1_1_FW", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/n8vubr98", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/forward/L6/400M/sae_weights.safetensors b/forward/L6/400M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..5d61681093393a83059e1d38b4b5048ce9cf6ee0
--- /dev/null
+++ b/forward/L6/400M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0dda6b54b6d8800d582a7566e2f4e396698380fbef34bc3c5019ede077f92990
+size 37801344
diff --git a/forward/L6/400M/sparsity.safetensors b/forward/L6/400M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..d6cb3b0e52fd6cc176e38984dec062b43b71cc99
--- /dev/null
+++ b/forward/L6/400M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d90f97065a9deaa4733acc73b7baae28922523fb211f8e9e002e31a996b1535c
+size 24656
diff --git a/forward/L6/500M/cfg.json b/forward/L6/500M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..16d9484eac4031b59a29953628b897a659a45bae
--- /dev/null
+++ b/forward/L6/500M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.6.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 6, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L5", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L6_hook_resid_post_L1_1_FW", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/n8vubr98", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/forward/L6/500M/sae_weights.safetensors b/forward/L6/500M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..e645b5ef03ffe6c31c4e2365d4cf9c355262d998
--- /dev/null
+++ b/forward/L6/500M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f7a2408f406bacfe5b432d4930ea90e6afc37991c40d3363de763e3de60155ff
+size 37801344
diff --git a/forward/L6/500M/sparsity.safetensors b/forward/L6/500M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..fc6cb6339b4901e85fd014eb29fa5be490de45fb
--- /dev/null
+++ b/forward/L6/500M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:186a2364d7bb809528b54d9d0ab6b16cb9e1ea71668b434848b1a5db436973e7
+size 24656
diff --git a/forward/L7/100M/cfg.json b/forward/L7/100M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..08c7292007f73463449d200e3d55905a220dbd65
--- /dev/null
+++ b/forward/L7/100M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.7.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 7, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L6", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L7_hook_resid_post_L1_1_FW", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/blne5ifx", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/forward/L7/100M/sae_weights.safetensors b/forward/L7/100M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..ef2ef18ecdb3ccf323e2b01b706b91b57a303ea4
--- /dev/null
+++ b/forward/L7/100M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:60af76fe6802592609864cd968e2bbd8dfcd115cef194b8e721526f4278969ac
+size 37801344
diff --git a/forward/L7/100M/sparsity.safetensors b/forward/L7/100M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..8257e513ad290079cb1f9a5e477c6569da2ec984
--- /dev/null
+++ b/forward/L7/100M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:1815b7b64de697f3b7fdb537d7d9a48657deaf14f2fd2e8c1cb2b1b9fb3bc87c
+size 24656
diff --git a/forward/L7/200M/cfg.json b/forward/L7/200M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..08c7292007f73463449d200e3d55905a220dbd65
--- /dev/null
+++ b/forward/L7/200M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.7.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 7, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L6", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L7_hook_resid_post_L1_1_FW", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/blne5ifx", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/forward/L7/200M/sae_weights.safetensors b/forward/L7/200M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..43da8e6f37bedfd44c2fb490004c27c759037cf1
--- /dev/null
+++ b/forward/L7/200M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f6631f2a17894881bef7119cffd9988721f8c2b75685554fe717d0f46c612074
+size 37801344
diff --git a/forward/L7/200M/sparsity.safetensors b/forward/L7/200M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..2c633a773e5596d013893b039c62cde600d180d0
--- /dev/null
+++ b/forward/L7/200M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fb5d6f7be6e09f1469e0a3f6b9804c70de1f0b26ff3486e9ea9f59efbbec98a0
+size 24656
diff --git a/forward/L7/300M/cfg.json b/forward/L7/300M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..08c7292007f73463449d200e3d55905a220dbd65
--- /dev/null
+++ b/forward/L7/300M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.7.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 7, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L6", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L7_hook_resid_post_L1_1_FW", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/blne5ifx", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/forward/L7/300M/sae_weights.safetensors b/forward/L7/300M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..fce561629333f12ce2ed5995eb63d20020773c6d
--- /dev/null
+++ b/forward/L7/300M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d9a396a5d5f2669a2d86b5d2fb977eddf5182d0b1717df2a32aa14758deddf09
+size 37801344
diff --git a/forward/L7/300M/sparsity.safetensors b/forward/L7/300M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..a40b6995e00f6f0bd1420ebd16c82aa0cd58bdab
--- /dev/null
+++ b/forward/L7/300M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:e5d9a841bf278ad3c54978256aaba031a4df5c2365006f9ba91bfd2c360a1036
+size 24656
diff --git a/forward/L7/400M/cfg.json b/forward/L7/400M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..08c7292007f73463449d200e3d55905a220dbd65
--- /dev/null
+++ b/forward/L7/400M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.7.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 7, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L6", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L7_hook_resid_post_L1_1_FW", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/blne5ifx", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/forward/L7/400M/sae_weights.safetensors b/forward/L7/400M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..846d2f800fc12c67d79843168d23bc481a9f410a
--- /dev/null
+++ b/forward/L7/400M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2b5a6162aa25f3ed51226eb899e0e9c31cbf11d35be4666b4a8acc9d01fb93fa
+size 37801344
diff --git a/forward/L7/400M/sparsity.safetensors b/forward/L7/400M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..be9765b61352b4acfdc5f9f4476d73e0fd9493e8
--- /dev/null
+++ b/forward/L7/400M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:72991c09f28ed2449ab395392fa6ba504f7571b28a3fc974752458ab90d5e585
+size 24656
diff --git a/forward/L7/500M/cfg.json b/forward/L7/500M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..08c7292007f73463449d200e3d55905a220dbd65
--- /dev/null
+++ b/forward/L7/500M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.7.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 7, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L6", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L7_hook_resid_post_L1_1_FW", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/blne5ifx", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/forward/L7/500M/sae_weights.safetensors b/forward/L7/500M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..2f1a26e53bb97e759a1f7104ee4d30e9e11c0036
--- /dev/null
+++ b/forward/L7/500M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2e9566556ec0e978dcdbc8f65402473c2b6b94f87d53ccdc1b1991509a106670
+size 37801344
diff --git a/forward/L7/500M/sparsity.safetensors b/forward/L7/500M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..7cafd6fdc2fa3eadd61a677f80f87e60411efcc8
--- /dev/null
+++ b/forward/L7/500M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:328caa01ae7c0ef81f7e6fca48e310e23a7aa05902bbf7853f28242f91e0b00a
+size 24656
diff --git a/forward/L8/100M/cfg.json b/forward/L8/100M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..80206666712f924748674e3c8d54397036955a44
--- /dev/null
+++ b/forward/L8/100M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.8.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 8, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L7", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L8_hook_resid_post_L1_1_FW", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/pd0rn85o", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/forward/L8/100M/sae_weights.safetensors b/forward/L8/100M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..dad431f895cfa7bf13bc92080de9edc1c50a95d9
--- /dev/null
+++ b/forward/L8/100M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:aa55ab505b367df6e81b32d47d88c55ea9263134cfae03740fd65b3aecd0137b
+size 37801344
diff --git a/forward/L8/100M/sparsity.safetensors b/forward/L8/100M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..88527da5bd0c3608d818286a157e8476fcf5b56a
--- /dev/null
+++ b/forward/L8/100M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:bd9e46d3920083bc948c02e724f692ed5891e7f6410d8c52de3b4ec814863f58
+size 24656
diff --git a/forward/L8/200M/cfg.json b/forward/L8/200M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..80206666712f924748674e3c8d54397036955a44
--- /dev/null
+++ b/forward/L8/200M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.8.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 8, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L7", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L8_hook_resid_post_L1_1_FW", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/pd0rn85o", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/forward/L8/200M/sae_weights.safetensors b/forward/L8/200M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..07a264de2ed3d58bb134b0fa58d63f0d19769935
--- /dev/null
+++ b/forward/L8/200M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:355e8b5f0c4bd1cfc4252a229d44476e7315aa78f3da4ed65343ec2cca5a8d12
+size 37801344
diff --git a/forward/L8/200M/sparsity.safetensors b/forward/L8/200M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..112384a6ad7703b81740b1a76ab72a873e60a0dd
--- /dev/null
+++ b/forward/L8/200M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:fe5088ea8e1e0d7be9046067e7bccad3f1de416edeeb77cabe7faccae3060ab0
+size 24656
diff --git a/forward/L8/300M/cfg.json b/forward/L8/300M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..80206666712f924748674e3c8d54397036955a44
--- /dev/null
+++ b/forward/L8/300M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.8.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 8, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L7", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L8_hook_resid_post_L1_1_FW", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/pd0rn85o", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/forward/L8/300M/sae_weights.safetensors b/forward/L8/300M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..7bee2c58a7f1c0a3652f583b9010a67c1d4416aa
--- /dev/null
+++ b/forward/L8/300M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:8fadfdbda325c1930383dcebcab591ecf1ed5613ef3d177ac8fe16dba68793b2
+size 37801344
diff --git a/forward/L8/300M/sparsity.safetensors b/forward/L8/300M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..8c2ae01c7aa4db6bdb3fe292fc9600447d54fe0e
--- /dev/null
+++ b/forward/L8/300M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:14ca30e38e9aea05b14015a7bf6c6ad51443b149f1f05d49c3d96d9fdc54cf40
+size 24656
diff --git a/forward/L8/400M/cfg.json b/forward/L8/400M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..80206666712f924748674e3c8d54397036955a44
--- /dev/null
+++ b/forward/L8/400M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.8.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 8, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L7", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L8_hook_resid_post_L1_1_FW", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/pd0rn85o", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/forward/L8/400M/sae_weights.safetensors b/forward/L8/400M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..a8fc2f24f7277d5f2f83df26f5818538ec34ed37
--- /dev/null
+++ b/forward/L8/400M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:666d60ae02c4abc2b5d0b4f77ab1f77900f5943774d8d525a6106926632bd978
+size 37801344
diff --git a/forward/L8/400M/sparsity.safetensors b/forward/L8/400M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..cbfe1b3684f9c44e91fa9b760303fe0411239cbd
--- /dev/null
+++ b/forward/L8/400M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a189da3db59f5bdc2e87a87d6936976ec16622e1d9ff3e37540d107e577032e7
+size 24656
diff --git a/forward/L8/500M/cfg.json b/forward/L8/500M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..80206666712f924748674e3c8d54397036955a44
--- /dev/null
+++ b/forward/L8/500M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.8.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 8, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L7", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L8_hook_resid_post_L1_1_FW", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/pd0rn85o", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/forward/L8/500M/sae_weights.safetensors b/forward/L8/500M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..29582c1ff5c1f750bc3ccccc6c137cc35a5d5edc
--- /dev/null
+++ b/forward/L8/500M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:83e1defbfafc2a1903c2c048b943c345e9370de79844803b2858c66a122d4547
+size 37801344
diff --git a/forward/L8/500M/sparsity.safetensors b/forward/L8/500M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..35f4219835d617e88ea8b9381636a8553a3bc604
--- /dev/null
+++ b/forward/L8/500M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0d5e4aafc100b704a477f087e992e85dba05104afd7701dcdfed632e0bb45402
+size 24656
diff --git a/forward/L9/100M/cfg.json b/forward/L9/100M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..8e404d8934ff60f878be8c39d638404e84a376f0
--- /dev/null
+++ b/forward/L9/100M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.9.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 9, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L8", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L9_hook_resid_post_L1_1_FW", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/hqc3zt9g", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/forward/L9/100M/sae_weights.safetensors b/forward/L9/100M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..f710d0bee867b81c5cefedfbc25d7f2842b83e9b
--- /dev/null
+++ b/forward/L9/100M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b5b08b2c356bec6aa20ab492c5028287992be95d64d81fa8ada457564297a532
+size 37801344
diff --git a/forward/L9/100M/sparsity.safetensors b/forward/L9/100M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..dcb0dfceb9db5149fa59e085f438b01090f2b810
--- /dev/null
+++ b/forward/L9/100M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:767c17dfbf49d749bd082a02669e7117dfb4bf48a19842e73654056c03f55b5d
+size 24656
diff --git a/forward/L9/200M/cfg.json b/forward/L9/200M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..8e404d8934ff60f878be8c39d638404e84a376f0
--- /dev/null
+++ b/forward/L9/200M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.9.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 9, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L8", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L9_hook_resid_post_L1_1_FW", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/hqc3zt9g", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/forward/L9/200M/sae_weights.safetensors b/forward/L9/200M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..e5641c6c2e036e4b350846c6c43ba09d97281f57
--- /dev/null
+++ b/forward/L9/200M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:dae968f9b55c1b7b729e4127ad1e1747a118ef5618a4d70211323617ff86058f
+size 37801344
diff --git a/forward/L9/200M/sparsity.safetensors b/forward/L9/200M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..b5caf63206fb03dfe190218a58c5fa0bee55641d
--- /dev/null
+++ b/forward/L9/200M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:0688c66678614a31e6dbf6629b4a95958d2416504439fda82e715629f4a51064
+size 24656
diff --git a/forward/L9/300M/cfg.json b/forward/L9/300M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..8e404d8934ff60f878be8c39d638404e84a376f0
--- /dev/null
+++ b/forward/L9/300M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.9.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 9, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L8", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L9_hook_resid_post_L1_1_FW", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/hqc3zt9g", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/forward/L9/300M/sae_weights.safetensors b/forward/L9/300M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..5543ebe4b5e01970822b19e61e2eeff5c51a4361
--- /dev/null
+++ b/forward/L9/300M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:899ae15318c19bbd14aa733e5bda550de582f6a0721d448a2fdcaed805ccaed1
+size 37801344
diff --git a/forward/L9/300M/sparsity.safetensors b/forward/L9/300M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..51defd27bee72b8e6dd0247e76ff231ed058c20c
--- /dev/null
+++ b/forward/L9/300M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d62f819f5d3970b166e39c3ab49cf7c2253644bda3246aec2adf6c42e4e04aec
+size 24656
diff --git a/forward/L9/400M/cfg.json b/forward/L9/400M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..8e404d8934ff60f878be8c39d638404e84a376f0
--- /dev/null
+++ b/forward/L9/400M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.9.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 9, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L8", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L9_hook_resid_post_L1_1_FW", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/hqc3zt9g", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/forward/L9/400M/sae_weights.safetensors b/forward/L9/400M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..9280b1310e150a844cfe83f8f92b7177de8ee9af
--- /dev/null
+++ b/forward/L9/400M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b7c6d8eecd265f58bfa02b4dded90af6f444e749f0ea7d16e2438fdacf629101
+size 37801344
diff --git a/forward/L9/400M/sparsity.safetensors b/forward/L9/400M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..676d549ae6374fddaa7bee2cb08da88d7e6f0b09
--- /dev/null
+++ b/forward/L9/400M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:766d7c7fc69c40342a1b1186e90f9ba744c0c044c5a3b966309e96c652283184
+size 24656
diff --git a/forward/L9/500M/cfg.json b/forward/L9/500M/cfg.json
new file mode 100644
index 0000000000000000000000000000000000000000..8e404d8934ff60f878be8c39d638404e84a376f0
--- /dev/null
+++ b/forward/L9/500M/cfg.json
@@ -0,0 +1 @@
+{"model_name": "pythia-160m-deduped", "model_class_name": "HookedTransformer", "hook_name": "blocks.9.hook_resid_post", "hook_eval": "NOT_IN_USE", "hook_layer": 9, "hook_head_index": null, "dataset_path": "NeelNanda/pile-small-tokenized-2b", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": true, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "jumprelu", "d_in": 768, "d_sae": 6144, "b_dec_init_method": "zeros", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": true, "noise_scale": 0.0, "from_pretrained_path": "./hub/models--mech-interp--pythia-160m-deduped-rs-post/snapshots/ad21dc2dd4070805dbeb842dcbfa14e6ad74e2b9/L8", "apply_b_dec_to_input": false, "decoder_orthogonal_init": false, "decoder_heuristic_init": false, "init_encoder_as_decoder_transpose": false, "n_batches_in_buffer": 128, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "none", "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": false, "autocast": false, "autocast_lm": false, "compile_llm": false, "llm_compilation_mode": null, "compile_sae": false, "sae_compilation_mode": null, "adam_beta1": 0, "adam_beta2": 0.999, "mse_loss_normalization": null, "l1_coefficient": 1, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": false, "l1_warm_up_steps": 6103, "lr": 1e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 1.0000000000000002e-06, "lr_decay_steps": 24414, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-06, "n_eval_batches": 10, "eval_batch_size_prompts": null, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-transfer-learning", "wandb_id": null, "run_name": "FT_L9_hook_resid_post_L1_1_FW", "wandb_entity": null, "wandb_log_frequency": 30, "eval_every_n_wandb_logs": 100, "resume": false, "n_checkpoints": 5, "checkpoint_path": "checkpoints/hqc3zt9g", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {}, "sae_lens_version": "3.14.0", "sae_lens_training_version": "3.14.0", "tokens_per_buffer": 536870912}
\ No newline at end of file
diff --git a/forward/L9/500M/sae_weights.safetensors b/forward/L9/500M/sae_weights.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..f9e1ee05ada05bef81bc1a63ff812ca073a9730e
--- /dev/null
+++ b/forward/L9/500M/sae_weights.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:ab2829ad179da34f589d5b83de29cc294df072bab7bf295275c2c1ebe755b9e1
+size 37801344
diff --git a/forward/L9/500M/sparsity.safetensors b/forward/L9/500M/sparsity.safetensors
new file mode 100644
index 0000000000000000000000000000000000000000..070a69deca884bd373872c1f09dec6e12842ff73
--- /dev/null
+++ b/forward/L9/500M/sparsity.safetensors
@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:58a54de19240d748ca5d0a0682c45064eb45ede1b9553d1109e66af7115cb83c
+size 24656