Upload SAE blocks.2.hook_resid_post

blocks.2.hook_resid_post/cfg.json

@@ -1 +1 @@
-{"model_name": "Qwen/Qwen2-0.5B", "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": "HuggingFaceFW/fineweb", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": false, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "standard", "d_in": 896, "d_sae": 7168, "b_dec_init_method": "geometric_median", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": false, "noise_scale": 0.0, "from_pretrained_path": null, "apply_b_dec_to_input": true, "decoder_orthogonal_init": false, "decoder_heuristic_init": true, "init_encoder_as_decoder_transpose": true, "n_batches_in_buffer": 64, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "expected_average_only_in", "seqpos_slice": [null], "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": true, "jumprelu_init_threshold": 0.001, "jumprelu_bandwidth": 0.001, "autocast": true, "autocast_lm": true, "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": 2.5, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": true, "l1_warm_up_steps": 15000, "lr": 2e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 2.0000000000000003e-06, "lr_decay_steps": 0, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-08, "n_eval_batches": 10, "eval_batch_size_prompts": 2, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-qwen2-0.5b", "wandb_id": null, "run_name": "sae-qwen-0.5b-layer-10-l1-2.5-cs", "wandb_entity": null, "wandb_log_frequency": 500, "eval_every_n_wandb_logs": 25, "resume": false, "n_checkpoints": 1, "checkpoint_path": "/cluster/project2/logiclm/sae-qwen2-500M/layer-10/ahkq3fxc", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {"center_writing_weights": false}, "sae_lens_version": "4.1.1", "sae_lens_training_version": "4.1.1", "tokens_per_buffer": 268435456}
+{"model_name": "Qwen/Qwen2-0.5B", "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": "HuggingFaceFW/fineweb", "dataset_trust_remote_code": true, "streaming": true, "is_dataset_tokenized": false, "context_size": 1024, "use_cached_activations": false, "cached_activations_path": null, "architecture": "standard", "d_in": 896, "d_sae": 7168, "b_dec_init_method": "geometric_median", "expansion_factor": 8, "activation_fn": "relu", "activation_fn_kwargs": {}, "normalize_sae_decoder": false, "noise_scale": 0.0, "from_pretrained_path": null, "apply_b_dec_to_input": true, "decoder_orthogonal_init": false, "decoder_heuristic_init": true, "init_encoder_as_decoder_transpose": true, "n_batches_in_buffer": 64, "training_tokens": 500000000, "finetuning_tokens": 0, "store_batch_size_prompts": 8, "train_batch_size_tokens": 4096, "normalize_activations": "expected_average_only_in", "seqpos_slice": [null], "device": "cuda", "act_store_device": "cuda", "seed": 42, "dtype": "float32", "prepend_bos": true, "jumprelu_init_threshold": 0.001, "jumprelu_bandwidth": 0.001, "autocast": true, "autocast_lm": true, "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": 2.5, "lp_norm": 1, "scale_sparsity_penalty_by_decoder_norm": true, "l1_warm_up_steps": 15000, "lr": 2e-05, "lr_scheduler_name": "constant", "lr_warm_up_steps": 0, "lr_end": 2.0000000000000003e-06, "lr_decay_steps": 0, "n_restart_cycles": 1, "finetuning_method": null, "use_ghost_grads": false, "feature_sampling_window": 2000, "dead_feature_window": 1000, "dead_feature_threshold": 1e-08, "n_eval_batches": 10, "eval_batch_size_prompts": 2, "log_to_wandb": true, "log_activations_store_to_wandb": false, "log_optimizer_state_to_wandb": false, "wandb_project": "sae-qwen2-0.5b", "wandb_id": null, "run_name": "sae-qwen-0.5b-layer-2-l1-2.5-cs", "wandb_entity": null, "wandb_log_frequency": 500, "eval_every_n_wandb_logs": 25, "resume": false, "n_checkpoints": 1, "checkpoint_path": "/cluster/project2/logiclm/sae-qwen2-500M/layer-2/r3eolnhl", "verbose": true, "model_kwargs": {}, "model_from_pretrained_kwargs": {"center_writing_weights": false}, "sae_lens_version": "4.1.1", "sae_lens_training_version": "4.1.1", "tokens_per_buffer": 268435456}

blocks.2.hook_resid_post/sae_weights.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:09d93c13984e0c3ec59ee1d8a6d4c34e115205e98959e735667a9a8ef706ac1d
+oid sha256:725955d6582947abf22d9bcb6b2f0d145f04006829ee05e0c86036ff028c9a6f
 size 51412792

blocks.2.hook_resid_post/sparsity.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:90745b81174c1b12ca89360b8b1c7ec8b727fb37fc34e681b33ebb4fb4873b97
+oid sha256:02fd310bfe55def30aaee08562e2ada5ef8abb8d847fc63dc5b0e5a8ae78641b
 size 28752