aicity2026-track3-cosmos3-super-sft

Merged LoRA SFT of nvidia/Cosmos3-Super — the cosmos_v2 model (all tasks, tiers S/A + Vad-R1, 32 frames) of the 2nd-place system for AI City Challenge 2026 Track 3 (traffic-anomaly video QA, nine-task test mean 0.6696). Derivative of Cosmos3-Super; inherits the OpenMDW 1.1 license of the base.

• Training recipe, data prep, and the full inference pipeline: https://github.com/HoangVuAn/aicity2026-track3-metric-matched-selection (this checkpoint = checkpoints/cosmos_sft_merged, step 3800 of the released recipe)
• Fine-tuned on the official AI City Challenge 2026 Track 3 annotations + the public Vad-Reasoning-SFT set. Training data is NOT redistributed here.
• Serve with vLLM (needs the cosmos plugins pinned in the repo's requirements.txt): vllm serve AnHoang200901/aicity2026-track3-cosmos3-super-sft --tensor-parallel-size 1

Paper: "Metric-Matched Test-Time Selection for Traffic Anomaly Video Question Answering", AI City Challenge Workshop @ ECCV 2026.