Description
我用如下脚本微调 qwen3-reranker-0.6B模型
NPROC_PER_NODE=2
swift sft
--model /data0/dikaer/jinyangyi/models/Qwen3-Reranker-0.6B
--dataset data/train_rerank.jsonl
--model_type qwen3
--torch_dtype bfloat16
--train_type full
--max_length 2048
--split_dataset_ratio 0
--output_dir output_qwen0.6B_rerank
--save_total_limit 1
--num_train_epochs 3
--logging_steps 100
--save_steps 500
--per_device_train_batch_size 32
--gradient_accumulation_steps 8
--learning_rate 8e-5
--loss_scale ignore_empty_think
--warmup_ratio 0.05 \
以下是我数据集的 jsonl 的格式
{"system": "Judge whether the Document meets the requirements based on the Query and the Instruct provided. Note that the answer can only be "yes" or "no".", "query": "3pcs Fierce Muscle Brown Bear Vinyl Decals - Durable Car Window Stickers for Glass & Metal Surfaces, Self-Adhesive Auto Accessories", "document": "Automotive-Exterior Accessories-Bumper Stickers, Decals & Magnets-Decals", "label": "yes"}
然后训练
{'loss': 0.0, 'grad_norm': 0.0, 'learning_rate': 1.9e-07, 'memory(GiB)': 34.06, 'train_speed(iter/s)': 0.226097, 'epoch': 0.0, 'global_step/max_steps': '1/8397', 'percentage': '0.01%', 'elapsed_time': '4s', 'remaining_time': '9h 23m 1s'}
loss 为0
请教一下是哪里出了问题