jxb1st/aicity2026-track3-7-8

AI City Challenge 2026 - Track 3 (TAR / PSI / FETV) LoRA adapters

LoRA adapters for our AI City 2026 Track 3 video-QA system. Code: https://github.com/jxb1st/aicity2026-track3-7-8

Subfolder	Base model	Track	Val
psi_cosmos7b	nvidia/Cosmos-Reason1-7B	PSI (Track 8)	~74
psi_qwen3vl32	Qwen/Qwen3-VL-32B-Instruct	PSI (Track 8)	70.5
psi_cosmos3nano	Cosmos3-Nano-VLM	PSI (Track 8)	72.45
psi_unified_qwen3vl32	Qwen/Qwen3-VL-32B-Instruct	Unified TAR+PSI	66.23
tar_qwen3vl8b_generalist	Qwen/Qwen3-VL-8B-Instruct	TAR (Track 3)	pipeline ~0.6144
tar_qwen3vl8b_temporal	Qwen/Qwen3-VL-8B-Instruct	TAR (Track 3)	temporal specialist
tar_qwen3vl8b_mcq	Qwen/Qwen3-VL-8B-Instruct	TAR (Track 3)	MCQ specialist
fetv_qwen3vl32_pseudo	Qwen/Qwen3-VL-32B-Instruct	FETV (Track 7)	-

The TAR specialists (tar_qwen3vl8b_*) stack on the merged generalist; see the repo tar/ README.