seed: 0
exp_name: pythia410m_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/pythia410m_hh_ga4_dpo_2023-07-20_13-33-11_249010
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-410m
  tokenizer_name_or_path: null
  archive: .cache/laura/pythia410m_hh_ga4_sft_2023-07-19_13-02-35_532776/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