Drop the no-op {{ bos_token }} from the chat template
#2
by qgallouedec HF Staff - opened
This PR drops the leading {{ bos_token }} from chat_template.jinja. bos_token is None for this tokenizer, so it renders as an empty string. Rendering is unchanged.
Without it, the template is byte-identical to the Qwen3 one, so tools that recognize Qwen3 pick it up. For example, TRL then swaps in its Qwen3 training template for assistant-only loss (SFT) and tool-calling training (GRPO). Today it raises, because it doesn't recognize the template.
Checked: the rendered prompt is identical before and after for 24 combinations (system / multi-turn / tool calls and responses, with and without tools, add_generation_prompt, enable_thinking).
qgallouedec changed pull request status to open
aakanksha19 changed pull request status to merged