ValueError：向已有特殊标记的分词器添加特殊标记；只有在打算添加新标记时才使用 `add_special_tokens=True`。收到 extra_ids=0，但分词器已有 2 个特殊标记。
ValueError: Adding special tokens to a tokenizer that already has them; use `add_special_tokens=True` only if you intend to add new tokens. Got extra_ids=0 but tokenizer already has 2 special tokens.
ID: huggingface/tokenizer-extra-special-tokens-invalid
版本兼容性
| 版本 | 状态 | 引入 | 弃用 | 备注 |
|---|---|---|---|---|
| tokenizers 0.19.1 | active | — | — | — |
| transformers 4.44.0 | active | — | — | — |
根因分析
用户使用空的或冗余的特殊标记字典调用了 `tokenizer.add_special_tokens()`，但分词器已定义这些标记，导致验证错误。
English
User called `tokenizer.add_special_tokens()` with an empty or redundant special tokens dictionary, but the tokenizer already has those tokens defined, causing a validation error.
官方文档
https://huggingface.co/docs/tokenizers/en/api/special-tokens解决方案
-
Check existing special tokens before adding: if `tokenizer.special_tokens_map` already contains the tokens, skip the `add_special_tokens` call entirely.
-
Use `tokenizer.add_special_tokens({'additional_special_tokens': ['<new_token>']})` only for truly new tokens, not duplicates.
无效尝试
常见但无效的做法:
-
100% 失败
The parameter `add_special_tokens` controls whether to add tokens to the vocabulary, not whether to check for duplicates; the error persists.
-
40% 失败
Deleting built-in special tokens (like [CLS], [SEP]) can break tokenizer functionality; re-adding may still fail if they are already present in the base tokenizer.