First model version

config.yaml

@@ -0,0 +1,46 @@
+seed: 0
+exp_name: pythia160m_hh_ga4_dpo
+batch_size: 32
+eval_batch_size: 32
+debug: false
+fsdp_port: 12355
+datasets:
+- hh
+wandb:
+  enabled: true
+  entity: pythia_dpo
+  project: Pythia_LOM
+local_dirs:
+- /scr-ssd
+- /scr
+- .cache
+sample_during_eval: false
+n_eval_model_samples: 16
+do_first_eval: true
+local_run_dir: .cache/laura/pythia160m_hh_ga4_dpo_2023-07-18_14-44-17_392406
+lr: 1.0e-06
+gradient_accumulation_steps: 4
+max_grad_norm: 10.0
+max_length: 512
+max_prompt_length: 256
+n_epochs: 1
+n_examples: null
+n_eval_examples: 256
+trainer: FSDPTrainer
+optimizer: RMSprop
+warmup_steps: 150
+activation_checkpointing: false
+eval_every: 20000
+minimum_log_interval_secs: 1.0
+model:
+  name_or_path: EleutherAI/pythia-160m
+  tokenizer_name_or_path: null
+  archive: .cache/laura/pythia160m_hh_ga4_sft_2023-07-18_13-09-00_785022/step-159744/policy.pt
+  block_name: GPTNeoXLayer
+  policy_dtype: float32
+  fsdp_policy_mp: null
+  reference_dtype: float16
+loss:
+  name: dpo
+  beta: 0.1
+  reference_free: false

optimizer.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:407dedde5609489597a721079189b955a983c72b5211be30f2ce81ce0d10135e
+size 649342149

policy.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:a20a36bff1a38bc01fdfda69c0ffe9bd95f996df06d515f31fe13388f4f78266
+size 699693271

scheduler.pt

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7e977d9cf46159b73094c47549966e67b87809ee469cc99b5ff372d0415c3fce
+size 627