Training in progress, step 25000

blimp_results.json

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:2def99d5596a716505ba5ff92e5d3c92b5656f6ea6bc185d219fe26776261522
-size 80215544
+oid sha256:f2147014be6c043e2b8ed41533487d0c39d674a7c078bb4d410bf8b3402612c4
+size 80218658

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_1725017275.3935974.yaml

@@ -0,0 +1,44 @@
+experiment:
+  seed: 42
+  name: gpt2_5M-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: 256
+    n_positions: 256
+    n_inner: 1024
+    resid_pdrop: 0.3
+    embd_pdrop: 0.3
+    attn_pdrop: 0.3
+trainer:
+  batch_size: 32
+  lr: 0.001
+  num_warmup_steps: 100000
+  max_training_steps: 600000
+  logging_steps: 6000
+  save_steps: 25000
+  eval_steps: 25000

model.safetensors

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:65a12d689bb9f43ee19c23d015355a4a4b2512d49ad3761b63f9c9fe5a595c14
+oid sha256:af76708dfc52079dd81be19bbde53a58516495d3b59f93b8dac1c20108155025
 size 35609928

training_args.bin

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:fc05a9b18c09d9343d851f614a906fd152098177d5c8db3b58ee8a0f0ce54453
+oid sha256:c9137eae5b2fc0ccaecfcc834c539253a235e58790318a23a66c11247c57014a
 size 5368