情感分类微调：测试输出出现大量positive和negative。

[un0o7]

Reminder

• I have read the README and searched the existing issues.

System Info

linux

Reproduction

instruction = f'Judge whether the movie review is positive or negative.'
fine_tune_data = [{
    'instruction': instruction,
    'input': data['text'],
    'output': 'positive' if data['label'] == 1 else 'negative'
} for data in train_data]

json.dump(fine_tune_data, open('LLaMA-Factory/data/imdb.json', 'w'), indent=2)

模型的输出有大量的negative和positive，并且lora合并之后，推理速度明显下降。

Expected behavior

仅输出一个positive或negative

Others

No response

[hiyouga]

llama3 非 instruct 模型请用 template: default 微调

[un0o7]

感谢大佬！

[nvliajia]

请问你微调分类任务有改动默认参数吗，为什么我微调后测试的时候发现有些答案不固定呢