huggingface config_error ai_generated true

用户警告：你正在使用一个仅解码器模型，且 padding_side='right'。这可能会产生错误结果。建议在分词前设置 padding_side='left'。

UserWarning: You are using a decoder-only model with padding_side='right'. This may produce incorrect results. Consider setting padding_side='left' before tokenizing.

ID: huggingface/decoder-only-padding-side-warning

其他格式: JSON · Markdown 中文 · English
92%修复率
88%置信度
1证据数
2023-06-15首次发现

版本兼容性

版本状态引入弃用备注
transformers>=4.30.0 active — — —
torch>=2.0.0 active — — —
Python>=3.8 active — — —

根因分析

仅解码器模型（如 GPT、LLaMA）期望在左侧进行填充以保持因果掩码；右侧填充会导致注意力机制在序列中间关注填充标记。

English

Decoder-only models (e.g., GPT, LLaMA) expect padding on the left side to maintain causal masking; right padding causes attention to attend to padding tokens in the middle of the sequence.

generic

官方文档

https://huggingface.co/docs/transformers/en/pad_truncation#padding-and-truncation-for-decoder-only-models

解决方案

  1. Set padding_side='left' and pad_token to eos_token before tokenizing:
from transformers import AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained('meta-llama/Llama-2-7b-chat-hf')
tokenizer.padding_side = 'left'
tokenizer.pad_token = tokenizer.eos_token
inputs = tokenizer(texts, padding=True, return_tensors='pt')
  2. Use a data collator that handles left padding automatically, such as DataCollatorForSeq2Seq with padding_side='left':
from transformers import DataCollatorForSeq2Seq
data_collator = DataCollatorForSeq2Seq(tokenizer, padding=True, pad_to_multiple_of=8)
tokenizer.padding_side = 'left'

无效尝试

常见但无效的做法:

  1. 70% 失败

    Setting padding_side='right' and adding an attention_mask is a common wrong fix; the model still sees padding in the middle of the sequence, breaking causal masking.

  2. 90% 失败

    Disabling padding entirely (padding=False) often causes batch processing to fail due to uneven sequence lengths, leading to a different error.