Add {% generation %} markers around assistant outputs Without these, apply_chat_template(return_assistant_tokens_mask=True) returns an all-zero mask, so answer_only_loss=True SFT packing silently falls back to all-token loss (now blocked by the megatron-bridge pack guard). Rendered text is byte-identical to the previous template on all conversation shapes; masks now match nemotron-think-tokenizer-prefill-parity exactly. Validated battery in GEOD-147.
#1
by Kyle1668 - opened
No description provided.
Kyle1668 changed pull request status to merged