QWEN-1.8B finetune 之后输出全是重复的 token
#1
by
xianf
- opened
我使用自己的 finetune 脚本进行 finetune，是通过 transformers 的 run_clm.py 进行数据处理上的简单修改，格式是 alpaca 的格式，数据的是 WMT 的翻译数据。输入的数据格式为：
# json 数据，由 instruction、input、output 组成
rank: 0 raw datasets: {
'instruction': 'Write a response that appropriately completes the request.\n\n### Request:\nTranslate from English to Chinese.\nThat would have meant too much pressure.\n\n### Response:',
'input': 'That would have meant too much pressure.',
'output': '这会给我带来太大的压力。',
'labels': '这会给我带来太大的压力。'
}
# -100 是 transformers 默认会忽略的 id，所以 label 用-100 进行填充
rank: 0 lm datasets: {
'labels': [-100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, 43288, 36993, 104169, 100393, 99222, 99750, 101950, 1773, 151643, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100, -100],
'input_ids': [7985, 264, 2033, 429, 34901, 44595, 279, 1681, 382, 14374, 6145, 510, 27473, 504, 6364, 311, 8453, 624, 4792, 1035, 614, 8791, 2238, 1753, 7262, 382, 14374, 5949, 25, 43288, 36993, 104169, 100393, 99222, 99750, 101950, 1773, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643, 151643],
'attention_mask': [1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]
}
# decode 得到的结果，格式符合预期
rank: 0 detok_input_ids: Write a response that appropriately completes the request.
### Request:
Translate from English to Chinese.
That would have meant too much pressure.
### Response:这会给我带来太大的压力。<|endoftext|><|endoftext|><|endoftext|>... #后面全是 endoftext
finetune 完的 loss 结果看起来比较正常：
***** train metrics *****
epoch = 1.0
train_loss = 1.7106
train_runtime = 0:04:27.49
train_samples = -1
train_samples_per_second = 169.825
train_steps_per_second = 0.265
[2023-12-05 11:54:55,808] [INFO] [launch.py:347:main] Process 260 exits successfully.
当模型进行 inference 的时候，生成的结果全是重复的字符：
input_ids: {'input_ids': tensor([[ 7985, 264, 2033, 429, 34901, 44595, 279, 1681, 382, 14374,
6145, 510, 27473, 279, 2701, 22870, 504, 6364, 311, 8453,
624, 2121, 807, 834, 1605, 432, 11, 422, 679, 3855,
311, 10955, 911, 6438, 1059, 6562, 13, 1036, 3838, 653,
358, 1977, 311, 1105, 30, 358, 1265, 19131, 7037, 11,
4658, 13, 3369, 13048, 0, 358, 2776, 422, 679, 3159,
7511, 14374, 5949, 25]], device='cuda:0'), 'token_type_ids': tensor([[0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0,
0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0]], device='cuda:0'), 'attention_mask': tensor([[1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1,
1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1, 1]], device='cuda:0')}
generated: Write a response that appropriately completes the request.### Request:Translate the following sentences from English to Chinese.As they neared it, Dory started to worry about meeting her parents. “What do I say to them? I should introduce myself, probably. ‘Hi! I'm Dory.'”### Response:::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::::
There was an update made for attention that fix potential bug in fine tuning, did you manage to get to the latest commit before your that round of fine tuning?
There was an update made for attention that fix potential bug in fine tuning, did you manage to get to the latest commit before your that round of fine tuning?
Thanks for your reply. I made an update with the latest commit and the output is still repeating some tokens but the repeated token has changed to some other tokens. Still trying to figure it out...
兄弟，你后面解决了没，我们现在也是这个问题，用的qwen-14B ，微调之后好多重复标签。
兄弟，你后面解决了没，我们现在也是这个问题，用的qwen-14B ，微调之后好多重复标签。
我当时的问题是使用了 label_smoothing，trainer 做了个判断，Qwen 不在 MODEL_FOR_CAUSAL_LM_MAPPING_NAMES.values() 里面，所以没有做 label 的位移，导致训练都在预测 token 本身。