VideoRLVR-Data · Verifiable video reasoning data for RLVR
Training and evaluation data accompanying
Video Models Can Reason with Verifiable Rewards.
VideoRLVR-Data is built for a stricter question:
Can a video model generate a full visual trajectory that is not only realistic, but also rule-consistent, executable, and automatically verifiable?
The dataset contains procedurally generated visual reasoning tasks from three domains: Maze, FlowFree, and Sokoban. Each sample pairs an initial visual state and task prompt with a ground-truth solution video. These domains are designed for reinforcement learning with verifiable rewards, where generated videos can be parsed and checked by symbolic rule-based evaluators.
What's in this repo
| Path | Description |
|---|---|
train/ | Training split for VideoRLVR supervised fine-tuning and RL initialization |
test/ | Held-out evaluation split generated with disjoint random seeds |
train/maze/ | Maze training videos and metadata |
train/flowfree/ | FlowFree training videos and metadata |
train/sokoban/ | Sokoban training videos and metadata |
test/maze/ | Maze test videos and metadata |
test/flowfree/ | FlowFree test videos and metadata |
test/sokoban/ | Sokoban test videos and metadata |
train/metadata_all.csv | Combined training metadata |
test/metadata_all.csv | Combined test metadata |
*/metadata_maze.csv | Maze-only metadata |
*/metadata_flowfree.csv | FlowFree-only metadata |
*/metadata_sokoban.csv | Sokoban-only metadata |
Dataset composition
| Domain | Train | Test | Description |
|---|---|---|---|
maze | 10,000 | 1,000 | Grid navigation from start to goal while avoiding walls |
flowfree | 10,000 | 1,000 | Connect colored endpoints with non-overlapping paths that fill the grid |
sokoban | 10,000 | 1,000 | Push boxes to target locations under Sokoban transition rules |
| Total | 30,000 | 3,000 | Multi-task verifiable video reasoning suite |
Each example contains an input state, a natural-language instruction, and a ground-truth solution video. Videos are rendered at 480×832 resolution with 81 frames.
How to use
# Download the dataset
hf download DarthZhu/VideoRLVR-Data \
--repo-type dataset \
--local-dir VideoRLVR-Data
cd VideoRLVR-Data
You can load the metadata files directly:
import pandas as pd
from pathlib import Path
root = Path("VideoRLVR-Data")
maze_train = pd.read_csv(root / "train" / "metadata_maze.csv")
flowfree_train = pd.read_csv(root / "train" / "metadata_flowfree.csv")
sokoban_train = pd.read_csv(root / "train" / "metadata_sokoban.csv")
Schema
The metadata files use a simple video-generation format.
{
"video": "relative/path/to/solution_video.mp4",
"input_image": "relative/path/to/input_image.png",
"prompt": "Natural-language task instruction"
}
Citation
@article{zhu2026video,
title={Video Models Can Reason with Verifiable Rewards},
author={Tinghui Zhu and Sheng Zhang and James Y. Huang and Selena Song and Xiaofei Wen and Yuankai Li and Hoifung Poon and Muhao Chen},
journal={arXiv preprint arXiv:2605.15458},
year={2026}
}