model_name_or_path: /data/mnpo/cache/qwen3_rev_smoke/transformers/models--Qwen--Qwen3-8B/snapshots/b968826d9c46dd6066d109eabc6255188de91218
torch_dtype: null
attn_implementation: eager
dataset_mixer:
  /data/mnpo/revision_qwen3_8b/full_iter1/precomputed/avg_oracle: 1.0
dataset_splits:
  - train
  - test
preprocessing_num_workers: 12
bf16: true
loss_type: simpo
dpo_beta: 0.05
simpo_beta: 2.5
simpo_gamma: 1.0
eta: 0.0075
ratio: 0.3333
max_history_t: 1
history_weights: [1.0]
beta: 10
learning_rate: 5.0e-7
lr_scheduler_type: cosine
warmup_ratio: 0.1
optim: adamw_torch
weight_decay: 0.0
seed: 42
gradient_accumulation_steps: 8
gradient_checkpointing: true
gradient_checkpointing_kwargs:
  use_reentrant: false
num_train_epochs: 1
per_device_train_batch_size: 2
per_device_eval_batch_size: 2
max_length: 2048
max_prompt_length: 1800
do_eval: true
eval_strategy: steps
eval_steps: 500
logging_steps: 10
log_level: info
generate_during_eval: false
metric_for_best_model: eval_loss
load_best_model_at_end: false
save_strategy: steps
save_steps: 1000
save_total_limit: 1
save_only_model: true
save_safetensors: true
push_to_hub: false
report_to:
  - wandb
output_dir: /data/mnpo/revision_qwen3_8b/full_iter1/train/simpo_avg_b2p5_g1p0_s42_h200
run_name: rev-q3-simpo-b2p5_g1p0-s42