---
license: other
license_name: source-dataset-terms
task_categories:
- video-text-to-text
- visual-question-answering
language:
- en
tags:
- arxiv:2609.25773
- video
- multi-hop-reasoning
- reinforcement-learning
- rlvr
- grpo
size_categories:
- 10K<n<100K
configs:
- config_name: default
  data_files:
  - split: train
    path: videohopchain_train.parquet
  - split: validation
    path: videohopchain_val.parquet
---

# Video-HopChain

Video-HopChain is a dataset of multi-hop video questions for reinforcement learning with verifiable
rewards (RLVR). Every question chains three to six yes/no sub-questions ("hops") about moments in one
video. Each hop carries two integers, of which the first counts when the answer to the hop is yes and
the second when the answer is no. The model adds the integers of all hops, and that sum is the answer,
so an exact match on one number gives a verifiable reward.

The integers of a question are resampled until every combination of hop answers gives a different sum.
A wrong intermediate hop therefore changes the total, which is what forces the chain to hold.

- **Paper:** *Video-HopChain: Multi-Hop Questions and Confidence-Gated Exploration for Video Reasoning Models* ([arXiv:2609.25773](https://arxiv.org/abs/2609.25773))
- **Model:** [ngqtrung/Video-HopChain-8B](https://huggingface.co/ngqtrung/Video-HopChain-8B)
- **Standard-RL checkpoint:** [ngqtrung/Video-HopChain-8B-Standard-RL](https://huggingface.co/ngqtrung/Video-HopChain-8B-Standard-RL)
- **Project page:** [ngquangtrung57.github.io/video-hopchain-page](https://ngquangtrung57.github.io/video-hopchain-page/)
- **Code:** [github.com/ngquangtrung57/video-hopchain](https://github.com/ngquangtrung57/video-hopchain)
- **Collection:** [Video-HopChain](https://huggingface.co/collections/ngqtrung/video-hopchain)

## Contents

| File | Rows | Size |
|---|---|---|
| `videohopchain_train.parquet` | 22,550 | 15.7 MB |
| `videohopchain_val.parquet` | 1,000 | 757 KB |
| `videohopchain_videos_longvila.zip` | — | 17.2 GiB |
| `videohopchain_videos_finevideo.zip` | — | 11.6 GiB |
| `videohopchain_videos_vript.zip` | — | 6.1 GiB |

The training split covers 13,378 videos. The held-out split holds 1,000 questions over 1,000 further
videos, and it is disjoint from the training split by video id. Every source video lasts at least
three minutes. The held-out split is stratified by hop count, by video duration and by source.

The video archives unpack to `videos/<source>/<id>.mp4`, which is the path that the `videos` field of
each row already points at. The clips are downscaled proxies, sampled at 140 frames and at most
50,176 pixels per frame, which is the resolution the questions were written and verified against.

## Row format

Each row follows the verl RLVR layout.

| Field | Meaning |
|---|---|
| `prompt` | chat messages, the system prompt and the question |
| `videos` | one entry with `video` (relative path), `fps`, `max_frames`, `max_pixels`, `min_pixels` |
| `data_source` | `hopchain_finevideo`, `hopchain_longvila` or `hopchain_vript` |
| `ability` | always `video_multihop_reasoning` |
| `reward_model` | `ground_truth` (the integer sum as a string) and `style` (`rule`) |
| `extra_info` | `answer`, `num_hops`, `hop_types`, `question_id`, `video_id`, `solve_rate`, `difficulty_band`, `tolerance` |

`tolerance` is `0.0`: the reward is an exact match on the sum.

## Usage

```python
from datasets import load_dataset

ds = load_dataset("ngqtrung/Video-HopChain")
print(ds["train"][0]["prompt"][-1]["content"])
print(ds["train"][0]["reward_model"]["ground_truth"])
```

Download and unpack the videos beside the parquets:

```bash
hf download ngqtrung/Video-HopChain --repo-type dataset --local-dir video-hopchain
cd video-hopchain && for z in videohopchain_videos_*.zip; do unzip -q "$z"; done
```

## How the questions were built

Source videos come from FineVideo, from LongVILA as released with the OneThinker training data, and
from the long split of Vript. Each video is cut into shots with PySceneDetect and every shot is
captioned once with MiniMax-M3. A language model then writes two candidate questions per video from
the captions and a sampled specification, and a judge loop checks each candidate against the captions
and rewrites it up to three times. The pipeline then ends with a difficulty filter that drops
questions the base model already solves in 6 of 8 rollouts. The code for all of this is in the
repository linked above.

Because the questions are written from captions rather than from the frames, a caption error can
reach a label. We report this limitation in the paper.

## Licensing

The questions and the annotations in the parquet files are released for research use.

The video archives re-distribute clips from FineVideo, LongVILA and Vript, as downscaled proxies.
Those clips remain under the terms of their original datasets and of the original uploaders, and this
repository does not grant any right beyond them. Check the source datasets before you use the video
for anything other than research, and contact us if you hold rights to a clip and want it removed.

## Citation

```bibtex
@misc{nguyenquang2026videohopchain,
  title         = {Video-HopChain: Multi-Hop Questions and Confidence-Gated Exploration for Video Reasoning Models},
  author        = {Nguyen Quang, Trung and Dong, Yuhao and Sun, Shuo and Liu, Shuai and Tian, Shulin and Yap, Kim-Hui and Liu, Ziwei},
  year          = {2026},
  eprint        = {2609.25773},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV},
  url           = {https://arxiv.org/abs/2609.25773}
}
```