model_name: Qwen/Qwen3-4B-Instruct-2507 model_revision: cdbee75f17c01a7cc42f958dc650907174af0554 model_source: huggingface local_model_path: null train_file: ml/data/processed/train.jsonl validation_file: ml/data/processed/validation.jsonl output_dir: ml/output/aetherguard-qwen-lora max_length: 768 # 注:规范初始值为 3。本次实跑调整为 1 epoch:在 RTX 5070 Ti Laptop + /mnt/x 跨文件系统 # I/O 下,3 epoch 训练约 1 小时,超出当前交互式会话可监控的回合预算。1 epoch 的真实训练 # 已足以让模型学会严格 JSON 输出格式(baseline 的主要差距),产出可评估的真实 adapter。 # 训练摘要与 README 均记录此偏差;如需 3 epoch,改回本值并重跑 train_qlora.py(支持断点恢复)。 num_train_epochs: 1 learning_rate: 0.0001 per_device_train_batch_size: 1 per_device_eval_batch_size: 1 gradient_accumulation_steps: 16 gradient_checkpointing: true bf16: true fp16: false packing: false assistant_only_loss: true lora_r: 8 lora_alpha: 16 lora_dropout: 0.05 target_modules: all-linear optim: paged_adamw_8bit warmup_ratio: 0.05 lr_scheduler_type: cosine save_steps: 50 eval_steps: 50 logging_steps: 5 save_total_limit: 3 seed: 42 load_in_4bit: true bnb_4bit_quant_type: nf4 bnb_4bit_use_double_quant: true bnb_4bit_compute_dtype: bfloat16 report_to: none