Training in progress, step 50000

blimp_results.json

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:e3c3e5231a5833ea81c3c627178853a917de1900b8646da982207cc3d00067e3
-size 80216800
+oid sha256:786c4ea1c8aee1d056e0da411594cb9fd8f38d1baa2aee83a9f5a8545e8a4d00
+size 80218105

config.json

@@ -3,9 +3,9 @@
   "architectures": [
     "GPT2LMHeadModel"
   ],
-  "attn_pdrop": 0.1,
+  "attn_pdrop": 0.3,
   "bos_token_id": 0,
-  "embd_pdrop": 0.1,
+  "embd_pdrop": 0.3,
   "eos_token_id": 0,
   "initializer_range": 0.02,
   "layer_norm_epsilon": 1e-05,
@@ -16,7 +16,7 @@
   "n_layer": 6,
   "n_positions": 256,
   "reorder_and_upcast_attn": false,
-  "resid_pdrop": 0.1,
+  "resid_pdrop": 0.3,
   "scale_attn_by_inverse_layer_idx": false,
   "scale_attn_weights": true,
   "summary_activation": null,

hydra_config_1724933085.355582.yaml

@@ -0,0 +1,44 @@
+experiment:
+  seed: 42
+  name: gpt2_19M-bpe-text-dyn-03
+  group: babylm-small
+  dry_run: false
+  offline_run: false
+  evaluate_segmentation: false
+  evaluate_babyslm: true
+  blimp_tasks: blimp_filtered,blimp_supplement
+  resume_checkpoint_path: null
+  resume_run_id: null
+dataset:
+  name: transformersegmentation/BabyLM-phonemized
+  subconfig: strict_small
+  text_column: text
+  is_phonemes: false
+  max_age: null
+tokenizer:
+  name: transformersegmentation/BabyLM-BPE-ortho-tokenizer
+data_preprocessing:
+  max_input_length: 128
+  join_utts: dynamic
+  remove_word_boundaries: false
+  subsample: null
+  subsample_type: examples
+model:
+  name: gpt2_lm
+  model_kwargs:
+    n_layer: 6
+    n_head: 8
+    n_embd: 512
+    n_positions: 256
+    n_inner: 2048
+    resid_pdrop: 0.3
+    embd_pdrop: 0.3
+    attn_pdrop: 0.3
+trainer:
+  batch_size: 32
+  lr: 0.001
+  num_warmup_steps: 90000
+  max_training_steps: 400000
+  logging_steps: 4000
+  save_steps: 50000
+  eval_steps: 50000

model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:8a1856b83481be0f7a52a705416d3ffb7c8d242e334e4ea6fdab8e61552617d9
+oid sha256:114a339d56e7961c6cae9800c50f675fe0f738602fa7e4cf307c1695c6030aff
 size 108961160

training_args.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:5dc80136f7f23f477ebdd661d182228c4de21c404425ed1d103d090bb1541237
+oid sha256:e0b5a8b6f1890201775a9dff957294eb0b004f3daffec4d13e5aebb84bcac43e
 size 5368