model_name_or_path: /NHNHOME/26msit001_A/BASE/AIPR/sjkim/revision_qwen3_8b/full_iter1/flagship_20260712/cache/huggingface/hub/models--Qwen--Qwen3-8B/snapshots/b968826d9c46dd6066d109eabc6255188de91218
torch_dtype: null
attn_implementation: sdpa
dataset_mixer:
  /NHNHOME/26msit001_A/BASE/AIPR/sjkim/revision_qwen3_8b/full_iter1/flagship_20260712/precomputed/avg: 1.0
dataset_splits:
- train
- test
preprocessing_num_workers: 4
bf16: true
loss_type: inpo
eta: 0.1
ratio: 0.3333
max_history_t: 1
history_weights:
- 1.0
dpo_beta: 0.05
simpo_beta: 2.0
simpo_gamma: 0.6
ronpo_alpha: 0.5
ronpo_tau: 0.05
ronpo_target_column: ronpo_target
reference_anchor_weight: 0.075
preference_sft_weight: 0.0075
ht_target_column: ht_target
ht_target_scale: 1.0
beta: 10.0
learning_rate: 5.0e-07
lr_scheduler_type: cosine
warmup_ratio: 0.2
optim: adamw_torch
weight_decay: 0.0
max_grad_norm: 1.0
seed: 42
gradient_accumulation_steps: 16
gradient_checkpointing: true
num_train_epochs: 1
max_steps: 900
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'
logging_steps: 5
log_level: info
generate_during_eval: false
load_best_model_at_end: false
save_strategy: steps
save_steps: 900
save_total_limit: 1
save_only_model: true
save_safetensors: true
push_to_hub: false
report_to:
- wandb
output_dir: /NHNHOME/26msit001_A/BASE/AIPR/sjkim/revision_qwen3_8b/full_iter1/fair_demo_20260715/sweep/candidates/inpo_b
run_name: qwen3-8b-fair-demo-inpo_b