One of the two specialists from Reason or Recite: Complementary Specialists with an LLM Judge for Traffic Anomaly Reasoning. Despite the repo name, this is the reasoning candidate: it produces a full reasoning trace before its answer. Its counterpart, ror-recite-27b, recites the scene description and summary as one fused answer, 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/Qwen3.5-9B, with a rank-64 LoRA merged in. These are the merged weights, so no adapter loading is needed.
Data: TAR-RE, the TAR (Traffic Anomaly Reasoning) training set with its reasoning traces regenerated by Qwen3.5-9B. Every gold answer is untouched; only the traces were rewritten.
Recipe: LoRA r=64, α=128, dropout 0.05, on the LM only; 4 epochs, AdamW 2e-4 on cosine with 3% warmup, bf16, max_length 5120, 32 frames at 131,072 px per frame.

How it is used in the submission

Candidate A answers all 960 test items via vLLM with greedy decoding (step 2 of the reproduction driver). The judge then compares its answer with the 27B's for each item; switching to this 9B requires unanimity across both candidate orders plus surviving a separate refutation pass.

hf download KakashiFromKonoha/ror-recite-9b --local-dir $REPRO_MODELS/reasoning-9b

Then follow the reproduction steps in the repo.

Reproducibility note

Greedy decoding through vLLM is not batch-invariant on this architecture (a Gated-DeltaNet hybrid; VLLM_BATCH_INVARIANT=1 is unsupported, see vllm#42960). Across cold passes a handful of items flip, which can move the judged submission by a few points. The exact leaderboard submission is checked into the repo at reproduction/leaderboard-winner.csv.