Add generation markers to the chat template
#2
by qgallouedec HF Staff - opened
Same as https://huggingface.co/allenai/AstaBrief_8B/discussions/3, applied to this checkpoint's template (stock Qwen3).
This PR adds {% generation %} / {% endgeneration %} markers around the assistant output in chat_template.jinja, so return_assistant_tokens_mask=True returns the assistant tokens, as needed for assistant-only loss (e.g. TRL SFT with assistant_only_loss=True). Rendering is unchanged. The mask covers what the model generates, up to and including <|im_end|>.
Looks good, thank you!
aakanksha19 changed pull request status to open
aakanksha19 changed pull request status to merged