Add response template parsing
#2
by nico-martin HF Staff - opened
Checkpoint-specific response-protocol audit.
This reasoning checkpoint uses <|cot_start|> / <|cot_end|>, not / . Keep thinking and content only. The message-end delimiter also closes a reasoning trace cut short at end of turn. Native tool calling is not documented for this math fine-tune.
Source: https://huggingface.co/LiquidAI/LFM2-350M-Math
response_template describes output parsing syntax, not a guarantee of tool-use quality. Existing chat_template formatting is unaffected.
Xenova changed pull request status to merged