OPD-TV: Rethinking Teacher Views for Privileged Supervision Transfer in Video On-Policy Distillation

GitHub code OPD-TV training data

Overview

This repository contains the OPD-TV Qwen3-VL-8B checkpoint used for the paper's main temporal-grounding evaluation.

OPD-TV studies a supervision transfer gap: enriching the teacher's temporal evidence improves its grounding accuracy, yet the resulting student gains are non-monotonic. The method compares the same teacher across the student's original visual input and an enriched teacher view, isolating the view-induced correction from the standard same-view policy correction. Token-level cross-view compatibility then regulates how strongly this correction enters a target anchored to the same-view teacher distribution.

The student generates trajectories from its original visual input. Privileged evidence is used for teacher supervision during training; inference uses the student alone.

Released artifacts

Artifact Link
Model checkpoint CewEhao/OPD-TV-Qwen3-VL-8B
Training data CewEhao/OPD-TV-Training-Data
Training and evaluation code Echochef/OPD-TV
Evaluation suite TencentARC/TimeLens-Bench

Evaluation

Scores are percentages.

Dataset mIoU R@0.3 R@0.5 R@0.7
Charades-STA 52.0 72.7 57.2 32.5
ActivityNet Captions 54.6 69.5 60.3 43.2
QVHighlights 67.6 82.1 73.5 58.3

The evaluation protocol uses uniform sampling at 2 FPS, a 768-frame cap, a 14,336 visual-token cap, greedy JSON decoding, and up to 64 generated tokens. See the repository instructions for the complete TimeLens-Bench setup.

Download and use

hf download CewEhao/OPD-TV-Qwen3-VL-8B \
  --local-dir models/OPD-TV-Qwen3-VL-8B

Load the checkpoint with Transformers or serve it with vLLM using --trust-remote-code.

After installing the code repository and preparing TimeLens-Bench, run its evaluation launcher:

export MODEL_PATH=CewEhao/OPD-TV-Qwen3-VL-8B
export BENCH_ROOT=/path/to/TimeLens-Bench
export OUTPUT_DIR=/path/to/opd-tv-eval
export CUDA_VISIBLE_DEVICES=0,1,2,3,4,5,6,7

bash scripts/eval_opd_tv_timelens.sh

License

The checkpoint is released under the Apache License 2.0.

Downloads last month
13
Safetensors
Model size
9B params
Tensor type
BF16
ยท
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for CewEhao/OPD-TV-Qwen3-VL-8B

Finetuned
(688)
this model

Dataset used to train CewEhao/OPD-TV-Qwen3-VL-8B