LoRA adapter for Qwen/Qwen3.6-27B, trained on 10% difficult-advice / 90% TULU3 by token count.

One arm of a mixture-ratio sweep. The 20/80 sibling cut ODCV-Bench misalignment from 37.2% → 19.2% against a matched FP8 base. The sweep asks whether the effect scales with the amount of difficult-advice data, plateaus, or tips into over-refusal — this is the low-dose arm.

All three hold total tokens at ~1.493M, so step count and compute are matched and the mixture ratio is the only variable.

Qwen3.6's chat template renders <think>{reasoning}</think> for any assistant turn that is final, so trace-free replay data would emit an empty <think></think> — the documented failure mode that trains a model to stop reasoning. The mixture builder appends a throwaway user turn to push the assistant off the end, hitting the template's no-think branch, then strips it. Verified on the written artifact: zero empty think blocks, think blocks in exactly the 147 difficult-advice rows.

bf16 LoRA (not QLoRA — bitsandbytes does not reliably cover this model's hybrid linear-attention/SSM layers), 1×H100 80GB, ~1h55m.

Packing is off because TRL only guarantees packed-sequence isolation under Flash Attention variants; under sdpa it warns of cross-contamination between samples. The vision tower (model.visual) is untouched.

Loss: 3.21 → ~0.96 by step 25, then flat (0.96–1.00). Final token accuracy 0.748, grad_norm 0.21, 1,487,733 tokens consumed.

Not yet evaluated on ODCV-Bench. When it is, the comparison will be against the same matched FP8 base arm (37.2%) used for the 20/80 result, on the same 78 scenario cells with the same two judges.

Use AutoModelForImageTextToText, not AutoModelForCausalLM — this is a vision-language checkpoint. Merging drops the base model's 15 mtp.* tensors, so speculative decoding needs them grafted back from the base checkpoint.