### 模型
model_name_or_path: Qwen/Qwen2.5-Coder-1.5B-Instruct
trust_remote_code: true

### 方法：LoRA 监督微调
stage: sft                    # 改成 dpo 就是 DPO 训练
do_train: true
finetuning_type: lora
lora_rank: 16                 # r
lora_alpha: 32                # 实际缩放 alpha / r = 2
lora_dropout: 0.05
lora_target: all              # 挂在所有线性层上：q/k/v/o + gate/up/down

### 数据
dataset: text2sql_train
template: qwen                # 对话模板，必须和模型匹配
cutoff_len: 1024              # 最长样本约 600 token，留足余量
overwrite_cache: true
preprocessing_num_workers: 4

### 输出
output_dir: saves/qwen2.5-coder-1.5b/lora/text2sql
logging_steps: 5
save_steps: 50
plot_loss: true               # 训练结束后自动画出损失曲线
overwrite_output_dir: true
report_to: none

### 训练
per_device_train_batch_size: 8
gradient_accumulation_steps: 2   # 等效 batch = 16
learning_rate: 2.0e-4            # LoRA 常用 1e-4 ~ 2e-4
num_train_epochs: 3.0
lr_scheduler_type: cosine
warmup_ratio: 0.05
max_grad_norm: 1.0
bf16: true

### 验证：盯着验证损失，而不是训练损失
val_size: 0.05
per_device_eval_batch_size: 8
eval_strategy: steps
eval_steps: 10