LoRA adapter trained on 498,035 tokens weighted toward maths (NuminaMath-CoT ~67%, TULU3 and No Robots ~16.5% each), with loss on assistant tokens only, for 1 epoch.

Training data: qwen3.6-27b-mixture-500k-numina-heavy.

Final loss 0.897 sits below the ~1M-token difficult-advice arms (0.88-0.93) at a comparable step count, and far below the 100k/14-step run (1.547). That reflects the data, not model quality: NuminaMath solutions are highly structured mathematical prose and an easier next-token target than open-ended advice. Do not read it as a better model.

Everything outside an assistant turn is -100. A supervised span starts immediately after the <|im_start|>assistant\n header and ends after the closing <|im_end|>, which the model must emit in order to stop. Verified: zero user or system tokens inside any supervised span.

TRL's own assistant_only_loss cannot do this on Qwen3.6 — its chat template has no {% generation %} markers. Spans are derived from the rendered text via the fast tokenizer's offset mapping instead, and TRL receives finished labels.

loss_type: nll is set because TRL's default chunked-CE path patches the LM head and reads forward.__func__, which fails on this checkpoint's functools.partial forward. The loss is mathematically the same.

Use AutoModelForImageTextToText, not AutoModelForCausalLM — this is a vision-language checkpoint.