minuzero

VideoKR-Qwen2.5-VL-7B

README

License: apache-2.0

About

This repository contains the VideoKR-Qwen2.5-VL-7B model presented in VideoKR: Towards Knowledge- and Reasoning-Intensive Video Understanding (ICML 2026 Spotlight).

VideoKR-Qwen2.5-VL-7B is obtained through a standard SFT → GRPO pipeline on Qwen2.5-VL-7B-Instruct:

Supervised fine-tuning on VideoKR-SFT-201K with CoT rationales → VideoKR-Qwen2.5-VL-7B-SFT
GRPO reinforcement learning on VideoKR-RL-114K with verifiable rewards → this model

VideoKR is the first large-scale training corpus designed for knowledge- and reasoning-intensive video understanding, containing 315K video reasoning examples over 145K newly collected, CC-licensed expert-domain videos across 82 professional subjects.

Links

Table with columns: Resource, Link
Resource	Link
Training data	minuzero/VideoKR-Train
Evaluation data	minuzero/VideoKR-Eval
SFT checkpoint (Qwen2.5-VL)	minuzero/VideoKR-Qwen2.5-VL-7B-SFT
SFT checkpoint (Qwen3-VL)	minuzero/VideoKR-Qwen3-VL-8B-SFT
GRPO checkpoint (Qwen3-VL)	minuzero/VideoKR-Qwen3-VL-8B

Performance

Results with 128 input frames. Within each base-model group, bold = best, underline = second best.

Table with columns: Model, Video-MME, MVBench, LongVBench, General Avg, VideoMMMU, MMVU, SciVidBench, VideoKR-Eval, Knowledge Avg
Model	Video-MME	MVBench	LongVBench	General Avg	VideoMMMU	MMVU	SciVidBench	VideoKR-Eval	Knowledge Avg
Qwen2.5-VL-7B-Instruct	65.1	66.3	60.9	64.1	51.1	55.7	28.1	32.7	41.9

VideoKR achieves the highest knowledge-intensive average (+4.7 over base, +2.3 over VideoAuto-R1) while remaining competitive on general video reasoning.

Results with 16 input frames (comparison with Video-R1 and VideoRFT):

Table with columns: Model, Video-MME, MVBench, LongVBench, General Avg, VideoMMMU, MMVU, SciVidBench, VideoKR-Eval, Knowledge Avg
Model	Video-MME	MVBench	LongVBench	General Avg	VideoMMMU	MMVU	SciVidBench	VideoKR-Eval	Knowledge Avg
Qwen2.5-VL-7B-Instruct	57.1	65.0	55.2	59.1	48.4	52.5	23.1	31.3	38.8

Under the 16-frame setting, VideoKR outperforms Video-R1 and VideoRFT by +4.2 and +4.0 on knowledge-intensive average, respectively.

Evaluation

bash
cd /path/to/VideoKR/lmms_eval
conda activate videokr_eval

export CUDA_VISIBLE_DEVICES=0
export VIDEOKR_MODEL=minuzero/VideoKR-Qwen2.5-VL-7B
export TASKS=videokr_eval
export BATCH_SIZE=1
export RUN_NAME=videokr_eval

bash examples/models/videokr_vllm.sh

Citation

If you find VideoKR useful in your research, please cite our paper:

bibtex
@misc{fu2026videokrknowledgereasoningintensivevideo,
      title={VideoKR: Towards Knowledge- and Reasoning-Intensive Video Understanding}, 
      author={Lin Fu and Zheyuan Yang and Yang Wang and Tingyu Song and Arman Cohan and Yilun Zhao},
      year={2026},
      eprint={2606.05259},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2606.05259}, 
}

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Container

Run this model inference with full control and performance in your environment.

Learn more

Model Details

Model Provider

minuzero

Model Tree

Base

Qwen/Qwen2.5-VL-7B-Instruct

Fine-tuned

this model

Input Modalities

TextImage

Output Modalities

Text

Supported Functionality

Dedicated EndpointsContainer

Explore FriendliAI today

Get started Talk to an engineer