loss_scale定制化失败

[jiusansan222]

不同的loss scale设置下微调qwen2.5-omni，模型在train loss上的表现与不自定义的loss scale表现一样，eval loss与不自定义的loss scale表现不一样，但是不同的scale表现还是一样的
脚本：

swift sft \
        --model Qwen2.5-Omni-7B \
        --dataset $train_data \
        --val_dataset $val_dataset \
        --system "You are a helpful assistant." \
        --train_type lora \
        --torch_dtype bfloat16 \
        --num_train_epochs 8 \
        --early_stop_interval 6 \
        --per_device_train_batch_size 4 \
        --per_device_eval_batch_size 8 \
        --learning_rate 1e-4 \
        --lora_rank 8 \
        --lora_alpha 32 \
        --target_modules all-linear \
        --freeze_vit true \
        --gradient_accumulation_steps $(expr 16 / $nproc_per_node) \
        --eval_steps 50 \
        --save_steps 50 \
        --save_total_limit 50 \
        --logging_steps 5 \
        --max_length 2048 \
        --output_dir $model_root \
        --warmup_ratio 0.05 \
        --dataloader_num_workers 8 \
        --external_plugins ms_swift_think_half/plugin_loss_scale_v3.py \
        --loss_scale think_half_v3 \

plugin_loss_scale_v3.py代码：

import os, json
from swift.plugin.loss_scale.loss_scale import LossScale, loss_scale_map

# Print on import so you can confirm the file is actually loaded
print("[swift] loading loss_scale plugin: think_half_v3")

class ThinkHalfV3(LossScale):
    # path to regex->weight json; set in __init__ for portability
    loss_scale_config = None

    def __init__(self, *args, **kwargs):
        super().__init__(*args, **kwargs)
        cfg = os.path.join(os.path.dirname(os.path.abspath(__file__)), "think_half_plus_empty_ignore.json")
        self.loss_scale_config = cfg
        try:
            type(self).loss_scale_config = cfg
        except Exception:
            pass

        # Print the patterns so you can verify the rules in logs
        try:
            with open(cfg, "r", encoding="utf-8") as f:
                data = json.load(f)
            print("[swift] think_half_v3 patterns loaded:")
            for k, v in data.items():
                print("   -", k, "=>", v)
        except Exception as e:
            print("[swift] think_half_v3 failed to read config:", e)

# Register the custom loss_scale into the global mapping
loss_scale_map["think_half_v3"] = ThinkHalfV3

think_half_plus_empty_ignore.json内容

{
  "<think>\\s*</think>": [0.0],                 
  "<think>[\\s\\S]*?</think>": [0.3],           
  "<answer>[\\s\\S]*?</answer>": [1.2]         
}

微调数据示例

{"messages": [{"role": "user", "content": "<audio> Please suggest XXX."}, {"role": "assistant", "content": "<think>\nExtract key XX balance\n</think>\n\n<answer>\n XXX.\n</answer>"}], "audios": ["4309.wav"]}
{"messages": [{"role": "user", "content": "<audio> Please XX  XXX."}, {"role": "assistant", "content": "<think>\nFor each XX  XXX\n</think>\n\n<answer>\nThe XX  XXX ment.\n</answer>"}], "audios": ["197.wav"]}
{"messages": [{"role": "user", "content": "<audio> Determine if this speech XX  XXX."}, {"role": "assistant", "content": "<think>\n</think>\n\n<answer>\nXX  XXX\n</answer>"}], "audios": ["SSB03090296.wav"]}

类似于上方的setting，设置了几个think的权重方案：紫红色为默认设置，绿色灰色为think权重0.2和0.5，橘黄色权重设置如上

观察训练的loss，train loss高度重合，eval loss三个一样，这是什么原因？请问是否脚本设置有误？