bear7011/streaming-vad-decision-adapters

Streaming VAD with decision models: LoRA adapters and derived weights

Weights behind Knows the Video, Not the Moment: Temporal Localisation Is the Missing Capability of Streaming Video Anomaly Detection (CVPR 2027 submission). Code, configs, per-window scores and the analysis scripts: https://github.com/Taide2026/system-one-streaming-vlm. This repo holds only the weights that are too large for git.

Layout

results/pro6000/<run name>/<time>/ mirrors the code repo's results folder. Each run folder holds

• adapter/ — peft LoRA (rank 64, alpha 128, every language-model projection of the Qwen3.5 backbone);
• readout.safetensors (d3: the 255-code read-out) or head.safetensors (Clef-Flash: the joint schema head), float32;
• calibration.json — the temperature fitted on held-out training videos; config.yaml, run.json — the training config.

Derived (zero-training) weights live directly under results/pro6000/<name>/: wise_a{25,5,75} (WiSE-FT: the hard-label adapter scaled toward the release), wisefov2_a{25,5,75} (the same on the foveated adapter), fovdial_a{25,5,75} (interpolation between fov2 and its hard-negative variant), dare* (DARE).

Run names

prefix	recipe
ft_ucf_v1_*, ft_xd_v1_*	single-task fine-tune on the V1 question (hard labels unless soft); _hard, _probe, _f4, _p10, _p30, _full, _hardneg are the ablations
floor_<model>_b	multi-task mix (50% V1 soft coverage, 25% 14-class, 25% text replay), no anchor
floor_<model>_c[_s1,_s2]	mix + KL anchor to the frozen release (the recipe of the paper), seeds
floor_<model>_d	the mix as DPO pairs
floor_<model>_c_hard, _b_hard	the mix with hard labels
floor_d3nano_mixv3	mix + KL with UCF101 clips as video replay
floor_d3nano_v1kl	V1 soft + KL only (no 14-class, no replay)
floor_d3nano_c_w3	mix + KL with the label-0 windows of anomalous training videos weighted x3 (localisation probe)
floor_xd_d3nano_c, floor_joint_d3nano_c	the recipe on XD-Violence; on UCF-Crime and XD-Violence jointly
fov_*, fov2_*, fov2h_*, fovcen_*, fovgate_*, fovfull2x_*	conditional foveation and its controls (unconditional, + hard negatives, centre crop, model gate, full-frame up-scaling)
fov2_xd_d3nano_c, fov2_d3flash_c, fov2_qwen9b_c, fov2_qwen2b_c	foveation on XD-Violence and on other backbones

Models: d3nano (d3-nano 2.2B), d3flash (8.4B), d327b (27B), clef (Clef-Flash 9B), qwen2b / qwen9b (plain Qwen3.5 controls).

Use

Point the code repo's eval config at a run folder: backbone: {name: d3, asset: d3-nano, adapter: results/pro6000/floor_d3nano_c/<time>}. The adapter is merged at load; inference is unchanged. Foveated runs need input_form: {name: foveated, frac: 0.333, upscale: 2, max_area: 0.045}.

Base models: d3-nano / d3-flash / d3 (vllm-sr), Clef-Flash (Cloudflare), Qwen3.5-2B / 9B (Apache-2.0). Training data: UCF-Crime and XD-Violence training videos with HIVAU-70k event times; no test video was read.