aicity2026-track3-f1-qwen3.6-27b
Merged LoRA SFT of Qwen/Qwen3.6-27B — the f1 model (primary open-ended generator, 128 frames,
7 BERTScore-scored tasks) of the 2nd-place system for AI City Challenge 2026 Track 3
(traffic-anomaly video QA, nine-task test mean 0.6696).
- Training recipe, data prep, and the full inference pipeline:
https://github.com/HoangVuAn/aicity2026-track3-metric-matched-selection
(this checkpoint =
checkpoints/qwen36_27b_f1_128_ck3600_merged, step 3600 of the released recipe)
- Fine-tuned on the official AI City Challenge 2026 Track 3 annotations (44,413-record corpus,
byte-reproducible via
data_prep/build_corpus.py in the repo). Training data is NOT redistributed here.
- Serve with vLLM:
vllm serve AnHoang200901/aicity2026-track3-f1-qwen3.6-27b --tensor-parallel-size 2
Paper: "Metric-Matched Test-Time Selection for Traffic Anomaly Video Question Answering",
AI City Challenge Workshop @ ECCV 2026.
