model_name_or_path: /path/to/Qwen2.5-7B-Instruct
trust_remote_code: false
stage: sft
do_train: true
do_eval: true
finetuning_type: lora
lora_rank: 16
lora_alpha: 32
lora_dropout: 0.05
lora_target: all
quantization_bit: 4
quantization_method: bnb
quantization_type: nf4
double_quantization: true
dataset_dir: ./data
dataset: train
eval_dataset: dev
template: qwen
cutoff_len: 2048
train_on_prompt: false
mask_history: true
preprocessing_num_workers: 4
overwrite_cache: true
output_dir: ./adapter
logging_steps: 1
save_strategy: epoch
eval_strategy: epoch
save_total_limit: 3
load_best_model_at_end: true
metric_for_best_model: eval_loss
greater_is_better: false
per_device_train_batch_size: 1
per_device_eval_batch_size: 1
gradient_accumulation_steps: 16
learning_rate: 0.0001
num_train_epochs: 3.0
lr_scheduler_type: cosine
warmup_ratio: 0.05
bf16: true
gradient_checkpointing: true
optim: paged_adamw_8bit
seed: 20261029
data_seed: 20261029
report_to: none
plot_loss: true
dataloader_num_workers: 2