### Qwen2.5-0.5B Full SFT Config
model_name_or_path: Qwen/Qwen2.5-0.5B-Instruct
stage: sft
do_train: true
finetuning_type: full
freeze_trainable_layers: 0

dataset: preference_extractor_train
template: qwen
cutoff_len: 1024
overwrite_cache: true
preprocessing_num_workers: 16

output_dir: saves/qwen2.5-0.5b-full-sft
logging_steps: 10
save_strategy: steps
save_steps: 500
plot_loss: true
overwrite_output_dir: true

per_device_train_batch_size: 16
gradient_accumulation_steps: 8
learning_rate: 2.0e-5
num_train_epochs: 1.0
lr_scheduler_type: cosine
warmup_ratio: 0.05
bf16: true
flash_attn: fa2

val_size: 0.01
per_device_eval_batch_size: 16
eval_strategy: steps
eval_steps: 500