model_name_or_path: /data/lry_machine_learning/models/Qwen2.5-7B-Instruct
stage: rm
do_train: true
finetuning_type: lora
lora_target: all
lora_rank: 16
lora_alpha: 32
lora_dropout: 0.05

dataset: ttc_rm
eval_dataset: ttc_rm_val
dataset_dir: data/training/rm
template: qwen
cutoff_len: 2048
overwrite_cache: false
preprocessing_num_workers: 8

output_dir: /data/lry_machine_learning/checkpoints/ttc_dense_verifier/verifier_qwen7b_rm_P10MULTI_20260604_055439
overwrite_output_dir: false
logging_dir: outputs/logs/verifier_rm_tensorboard_P10MULTI_20260604_055439
logging_steps: 10
eval_strategy: steps
save_strategy: steps
save_steps: 200
eval_steps: 200
save_total_limit: 3

per_device_train_batch_size: 1
per_device_eval_batch_size: 1
gradient_accumulation_steps: 6
learning_rate: 1.0e-5
num_train_epochs: 3.0
lr_scheduler_type: cosine
warmup_ratio: 0.03
bf16: true
ddp_timeout: 180000000
plot_loss: true

metric_for_best_model: eval_loss
load_best_model_at_end: false
greater_is_better: false