One of the two specialists from Reason or Recite: Complementary Specialists with an LLM Judge for Traffic Anomaly Reasoning. This is the recite candidate: it drops the reasoning trace entirely and is trained on answers alone. Its counterpart, ror-recite-9b, reasons before answering, and a Qwen3.6-27B judge picks between the two per item.

Code, configs, and the full submission driver: https://github.com/mayur-ag/reason-or-recite

What this checkpoint is

Base: Qwen/Qwen3.6-27B, with a rank-64 LoRA merged in. These are the merged weights, so no adapter loading is needed.
Data: the TAR (Traffic Anomaly Reasoning) training set, answers only; the adapter drops the reasoning field. scene_description and video_summarization are fused into a single <scene>…</scene>\n<summary>…</summary> target, so scene_description is excluded as a separate task.
Recipe: LoRA r=64, α=128, dropout 0.05 on the LM over an NF4-quantized base (vision tower kept in bf16); 2 epochs, AdamW 2e-4 on cosine with 3% warmup, 24 frames per clip. QLoRA is what fits the 27B run on a single GPU.

How it is used in the submission

Candidate B answers 880 test items via HF generate at batch size 1 (step 3 of the reproduction driver: the 80 scene_description items are covered by the fused answer, which step 4 splits back into scene and summary and from which mcq_openended is derived). The judge defaults to this candidate: switching to the 9B requires a unanimous verdict across both candidate orders plus a separate refutation pass.

hf download KakashiFromKonoha/ror-recite-27b --local-dir $REPRO_MODELS/qwen36-27b-recite-merged

Then follow the reproduction steps in the repo.