THUDM / ChatGLM-6B

ChatGLM-6B: An Open Bilingual Dialogue Language Model | 开源双语对话语言模型
Apache License 2.0
40.47k stars 5.19k forks source link

[BUG/Help] 推理得到的预测值全为nan #1400

Open nanisnull opened 1 year ago

nanisnull commented 1 year ago

Is there an existing issue for this?

Current Behavior

在使用chatGLM-6B-PT进行一个只有6条数据集的微调时,在推理阶段预测值全为0。请问这是因为模型参数设置有问题还是数据集设置有问题呢。

这是results.json { "epoch": 2400.0, "predict_bleu-4": NaN, "predict_rouge-1": NaN, "predict_rouge-2": NaN, "predict_rouge-l": NaN, "predict_runtime": 1.4923, "predict_samples": 6, "predict_samples_per_second": 4.021, "predict_steps_per_second": 4.021, "train_loss": 0.0006010692119598389, "train_runtime": 9556.3384, "train_samples": 6, "train_samples_per_second": 5.023, "train_steps_per_second": 0.314 }

这是生成预测文本: {"labels": "简约而不简单的牛仔外套,白色的衣身十分百搭。衣身多处有做旧破洞设计,打破单调乏味,增加一丝造型看点。衣身后背处有趣味刺绣装饰,丰富层次感,彰显别样时尚。", "predict": ""} {"labels": "你好,我是菜菜", "predict": ""} {"labels": "你好,我是菜菜,一个帅气的偶像。", "predict": ""} {"labels": "菜菜是一个帅气的偶像,擅长唱,跳,篮球和rap。最出名的歌曲是及你太美。", "predict": ""} {"labels": "菜菜是一个帅气的偶像,擅长唱,跳,篮球和rap。最出名的歌曲是及你太美。", "predict": ""} {"labels": "菜菜是一个帅气的偶像,擅长唱,跳,篮球和rap。最出名的歌曲是及你太美。", "predict": ""}

这是训练和测试文本: [{"content": "类型#上衣材质#牛仔布颜色#白色风格#简约图案#刺绣衣样式#外套衣款式#破洞", "summary": "简约而不简单的牛仔外套,白色的衣身十分百搭。衣身多处有做旧破洞设计,打破单调乏味,增加一丝造型看点。衣身后背处有趣味刺绣装饰,丰富层次感,彰显别样时尚。"} ,{"content": "你好,你是谁", "summary": "你好,我是菜菜"} ,{"content": "你是谁", "summary": "你好,我是菜菜,一个帅气的偶像。"} ,{"content": "菜菜是谁", "summary": "菜菜是一个帅气的偶像,擅长唱,跳,篮球和rap。最出名的歌曲是及你太美。"} ,{"content": "介绍下菜菜", "summary": "菜菜是一个帅气的偶像,擅长唱,跳,篮球和rap。最出名的歌曲是及你太美。"} ,{"content": "菜菜", "summary": "菜菜是一个帅气的偶像,擅长唱,跳,篮球和rap。最出名的歌曲是及你太美。"}]

Expected Behavior

No response

Steps To Reproduce

是根据repo的readme的操作的,修改的是训练和测试数据集

Environment

- OS:windows11
- Python:3.9
- Transformers:4.27.1
- PyTorch:1.12.1
- CUDA Support (`python -c "import torch; print(torch.cuda.is_available())"`) :True

Anything else?

No response