### model
model_name_or_path: F:\微调实验\llamafactory\models\Qwen\Qwen3-8B  # 或具体路径

### method
stage: sft
do_train: true
finetuning_type: lora
lora_target: all
quantization_bit: 4  # 关键！4bit量化

### dataset  
dataset: sft_complete_v3  # 用你之前的60条数据
template: qwen  
cutoff_len: 2048

### output
output_dir: saves/qwen3_8b_soviet_v2
overwrite_cache: true

### train
per_device_train_batch_size: 1  # 8B模型，保守点
gradient_accumulation_steps: 4
learning_rate: 3e-4
num_train_epochs: 30  # 8B不容易过拟合，可以少点
lora_rank: 64  # 8B模型可以用小点的rank
lora_alpha: 32

warmup_ratio: 0.1        # 加个warmup
gradient_checkpointing: true  # 省显存，能让batch更大