File size: 5,234 Bytes
0f7ec79 1728be3 0f7ec79 241a06c 0f7ec79 042c537 0f7ec79 2054d52 0f7ec79 cb72ded 0f7ec79 2054d52 0f7ec79 2054d52 0f7ec79 2054d52 0f7ec79 2054d52 0f7ec79 2054d52 0f7ec79 2054d52 0f7ec79 c671a0f ca7c887 0f7ec79 c90c7d4 0f7ec79 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 | ---
license: apache-2.0
task_categories:
- video-text-to-text
- visual-question-answering
language:
- en
tags:
- video
- long-video
- reasoning
- tool-calling
- multimodal
size_categories:
- 100K<n<1M
viewer: false
---
# ParaVT-Source
Source media archives for the [ParaVT](https://github.com/EvolvingLMMs-Lab/ParaVT) training corpus. Pair this repository with the annotations in [`ParaVT/ParaVT-Parquet`](https://huggingface.co/datasets/ParaVT/ParaVT-Parquet).
## Overview
ParaVT is a multi-agent agentic framework for long-video understanding, post-trained with **PARA-GRPO** (Parseability-Anchored and Ratio-gAted GRPO). This dataset bundles the raw video files referenced by every row in `ParaVT-Parquet`, packaged as per-source zip archives.
## Layout
Files are grouped by sentinel bucket. Each archive's members are stored under their full sentinel-form path, so extracting every zip into a single root produces a unified tree that [`paravt.data.materialize`](https://github.com/EvolvingLMMs-Lab/ParaVT/blob/main/paravt/data/materialize.py) can re-link in one call.
| Bucket | Contents | Archive convention |
|---|---|---|
| `longvt_source/<src>/` | LongVT shared training clips (`videor1_*`, `longvideoreason_*`, `geminicot_*`, `tvg_*`, `selftrace_*`) | `<src>_<idx>.zip`, mirroring the [`longvideotool/LongVT-Source`](https://huggingface.co/datasets/longvideotool/LongVT-Source) naming |
| `longvt_source/videor1_<N>/...` *(images)* | Auxiliary image files (`.png` / `.jpg`) referenced by the multi-modal interleaved rows in `paravt_sft_videor1_50k.parquet` | `videor1_<N>_images.zip` (one per `videor1_<N>` bucket) |
| `museg/charades/` | Charades-STA clips used by the `charades` SFT split and the `charades_tvg` RL split | `charades_<idx>.zip` |
| `museg/et_instruct_164k/` | MuSeG `et_instruct_164k` clips used by the `museg` SFT split | `et_instruct_<idx>.zip` |
| `selfqa/` | Self-curated open-ended QA clips (HACS- and Ego4D-derived UUIDs / YouTube IDs) used by the `hacs` and `ego4d_naq` RL splits | `selfqa.zip` (single archive, ~3 GB) |
Each archive is sized to stay below 10 GB on disk so that LFS pointer + Cloudflare CDN serving stays well-behaved. The `<src>_<idx>.zip` and `videor1_<N>_images.zip` archives that share a `videor1_<N>` prefix unzip into disjoint subdirectories of the same `videor1_<N>/` tree (videos under `CLEVRER/`, `LLaVA-Video-178K/`, … and images under `Chart/`, `Math/`, `Knowledge/`, `OCR/`, …), so they don't overwrite each other.
## Usage
```bash
# 1. Download every archive (use --include for a subset; see below).
huggingface-cli download ParaVT/ParaVT-Source --repo-type dataset --local-dir ./paravt-source
# 2. Extract every zip into the same root. Each zip's arcname carries the
# full sentinel path (e.g. "longvt_source/videor1_7/Math/...png"), so the
# extraction target must be the top-level root, NOT the per-zip directory.
( cd ./paravt-source && find . -name "*.zip" -exec unzip -q -o -d . {} \; )
# 3. Re-link absolute paths inside the parquets (one shot; see ParaVT/ParaVT-Parquet).
python -m paravt.data.materialize \
--root ./paravt-source \
--parquet-dir ./paravt-parquet \
--output-dir ./paravt-parquet-materialized
# Selective: pull only the buckets you need (e.g. Charades grounding only).
# Materialize will warn on missing files but produce valid output for the
# subset that is present.
huggingface-cli download ParaVT/ParaVT-Source \
--repo-type dataset --local-dir ./paravt-source \
--include "museg/charades.zip"
```
After step 2 the directory tree under `./paravt-source/` looks like:
```
paravt-source/
├── longvt_source/
│ ├── videor1_<N>/ # mp4 from videor1_<N>.zip + png/jpg from videor1_<N>_images.zip
│ │ ├── CLEVRER/, LLaVA-Video-178K/, NeXT-QA/, ... # videos
│ │ └── Chart/, Math/, Knowledge/, OCR/, ... # images
│ ├── longvideoreason_<N>/ # from longvideoreason_<N>_part{1,2}.zip
│ ├── geminicot_<N>/, tvg_<N>/, selftrace_<N>/, ...
├── museg/
│ ├── charades/ # mp4 from museg/charades.zip
│ └── et_instruct_164k/ # mp4 from museg/et_instruct.zip
└── selfqa/ # mp4 from selfqa/selfqa.zip
```
The materialized parquets then point at `file://<absolute-path-to-paravt-source>/<sentinel-path>` for every row.
## Citation
```bibtex
@misc{yang2026paravt,
title={{ParaVT}: Taming the Tool Prior Paradox for Parallel Tool Use in Agentic Video Reinforcement Learning},
author={Zuhao Yang and Kaichen Zhang and Sudong Wang and Keming Wu and Zhongyu Yang and Bo Li and Xiaojuan Qi and Shijian Lu and Xingxuan Li and Lidong Bing},
year={2026},
eprint={2605.20342},
archivePrefix={arXiv},
primaryClass={cs.CV}
}
```
## Acknowledgements
The `longvt_source/` archives reuse subsets of the [LongVT](https://github.com/EvolvingLMMs-Lab/LongVT) training media released at [`longvideotool/LongVT-Source`](https://huggingface.co/datasets/longvideotool/LongVT-Source); the MuSeG, Charades-STA, HACS, and Ego4D source clips are attributed to their respective original publications and used under their original licenses.
|