<think> token
#2
by
Downtown-Case
- opened
In the future, wouldn't it be better to add (or keep) the token in the tokenizer config? Right now it only has <|think|> (with no equivalent to ), and it seems your model was trained to use going by the eval.
If it's based on QwQ, doubly so, as then it can "inherit" the thinking delineator.
Yes! Our model starting with v26 has been switched to <think>, although we have not allocated a dedicated token for now.