model_name_or_path: /NHNHOME/WORKSPACE/26msit001_A/BASE/aipr_lab_sjkim_eval/revision_qwen3_8b/full_iter1/cache/huggingface/hub/models--Qwen--Qwen3-8B/snapshots/b968826d9c46dd6066d109eabc6255188de91218
torch_dtype: null
attn_implementation: eager
dataset_mixer:
  /NHNHOME/WORKSPACE/26msit001_A/BASE/aipr_lab_sjkim_eval/revision_qwen3_8b/full_iter1/precomputed/full_expect_bestadv_sample1: 1.0
dataset_splits:
  - train
  - test
preprocessing_num_workers: 12
bf16: true
loss_type: ronpo
max_history_t: 1
history_weights: [1.0]
ronpo_alpha: 1.0
ronpo_tau: 0.05
ronpo_target_column: ronpo_target
reference_anchor_weight: 0
preference_sft_weight: 0.005
beta: 10
learning_rate: 5.0000000000e-07
lr_scheduler_type: cosine
warmup_ratio: 0.1
optim: adamw_torch
weight_decay: 0.0
max_grad_norm: 1.0
seed: 42
gradient_accumulation_steps: 16
gradient_checkpointing: true
gradient_checkpointing_kwargs:
  use_reentrant: false
num_train_epochs: 1
max_steps: 75
per_device_train_batch_size: 1
per_device_eval_batch_size: 1
max_length: 2048
max_prompt_length: 1800
do_eval: false
eval_strategy: "no"
eval_steps: 500
logging_steps: 5
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: /NHNHOME/WORKSPACE/26msit001_A/BASE/aipr_lab_sjkim_eval/revision_qwen3_8b/full_iter1/train/ronpo_full_expect_s42_ronpo-bestadv-20260712-0750-r3-sample1_lr5e7_a1_s75_sft5e3
run_name: rev-q3-ronpo-full_expect-bestadv_sample1-s42-ronpo-bestadv-20260712-0750-r3-sample1_lr5e7_a1_s75_sft5e3