LoRA adapter trained on 499,595 tokens weighted toward maths (NuminaMath-CoT ~67%, TULU3 and No Robots ~16.5% each), with loss on assistant tokens only and the empty-think markers excluded from the loss, for 1 epoch.

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

The TULU3 (119) and No Robots (271) rows each carry <think>\n\n</think>\n\n at the start of their final assistant turn -- Qwen3.6's non-thinking marker, placed exactly where apply_chat_template puts it (asserted against the template before any data was touched). The 611 NuminaMath-CoT rows carry no marker: their chain-of-thought lives in the response text, so marking them "non-thinking" would contradict their own content.

Those marker tokens are excluded from the loss. The model is conditioned on the marker -- which is how Qwen3.6 injects it as a prefill in non-thinking mode -- but is never trained to emit one, since learning to emit an empty think block is the documented reasoning-collapse pattern.

<|im_start|>   MASKED
assistant      MASKED
<think>        MASKED   <- marker: context, not a target
</think>       MASKED
A              LOSS     <- supervision starts at the real answer

The arithmetic confirms it: 390,921 supervised tokens, identical to the marker-free build of the same mixture. Adding 390 markers x 4 tokens changed the supervised count by zero.

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.

Two variables differ between these (learning rate and the markers), so the pair does not isolate either. A marker-free run at 4e-5 would be needed for that.

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