videorl-video-rl-train

105,993-row multiple-choice video QA training set for GRPO/RLVR fine-tuning of vision-language models. Derived from five public video understanding benchmarks, re-curated at 24 frames / 100k pixels for throughput-efficient RL rollout. Used to train the Qwen3-VL-8B video RL models reported in the multimodal-rlvr project.

VIDEOS ARE NOT INCLUDED. Parquets store an external file:// path reference to the source video. You must download the original public datasets and place the videos under data/raw/ as described below.

---

Source breakdown

Config / parquet	data_source value	Rows	Public dataset
video_r1_llava_mc.parquet	video_r1_llava_mc	72,421	LLaVA-Video (Video-R1 split)
video_r1_nextqa_mc.parquet	video_r1_nextqa_mc	7,549	NExT-QA (Video-R1 split)
video_r1_perceptiontest_mc.parquet	video_r1_perceptiontest_mc	6,348	PerceptionTest (Video-R1 split)
video_r1_clevrer_mc.parquet	video_r1_clevrer_mc	8,220	CLEVRER (Video-R1 split)
video_r1_star_mc.parquet	video_r1_star_mc	11,455	STAR (Video-R1 split)
Total		105,993

All rows are multiple-choice with single-letter ground truths (A/B/C/D/E).

---

Frame and pixel configuration

This is the 24f100k variant — re-curated from the original Video-R1 parquets specifically for RL throughput:

Parameter	Value
fps	1
max_frames	24
min_pixels	3,136 (56×56)
max_pixels	100,000

The original 32-frame / 501,760-pixel variant exists as *_v1.parquet (without the _24f100k suffix). The 24-frame variant produces ~4x shorter prompt token sequences, substantially reducing rollout wall-time during GRPO training.

---

Schema

Each row has the following columns:

Column	Type	Description
prompt	list[{role: str, content: str}]	Chat-formatted prompt. First message is system (reasoning format instruction). Second is user (contains <video> placeholder + MCQ question).
images	list[{bytes, path}]	Always empty for video rows.
videos	list[{type, video, fps, max_frames, min_pixels, max_pixels}]	One entry per row. video is an external file:// path — not embedded bytes.
data_source	str	Source identifier, e.g. video_r1_llava_mc.
ability	str	Always video_understanding.
reward_model	{ground_truth: str, style: str}	ground_truth is a single letter (A–E). style is rule.
extra_info	{reward_type: str, answer: str}	reward_type is multiple_choice. answer mirrors ground_truth.

Prompt system message (baked in)

Every row embeds this reasoning system prompt verbatim:

You are a careful reasoning assistant. ALWAYS respond in this EXACT format:

<think>step-by-step reasoning</think>
<answer>\boxed{final_answer}</answer>

Example row (abbreviated)

{
  "prompt": [
    {"role": "system", "content": "You are a careful reasoning assistant. ALWAYS respond in this EXACT format:\n\n<think>step-by-step reasoning</think>\n<answer>\\boxed{final_answer}</answer>\n..."},
    {"role": "user",  "content": "<video>\nWhich of the following best describes what happens in the video?\nA. ...\nB. ...\nC. ...\nD. ...\nPlease reason step by step, then provide your final answer."}
  ],
  "images": [],
  "videos": [
    {
      "type": "video",
      "video": "file:///YOUR_DATA_ROOT/video-r1-extracted/LLaVA-Video-178K/liwei_youtube_videos/videos/youtube_video_2024/ytb_7nRmsEw7nsE.mp4",
      "fps": 1,
      "max_frames": 24,
      "min_pixels": 3136,
      "max_pixels": 100000
    }
  ],
  "data_source": "video_r1_llava_mc",
  "ability": "video_understanding",
  "reward_model": {"ground_truth": "A", "style": "rule"},
  "extra_info": {"reward_type": "multiple_choice", "answer": "A"}
}

---

Videos are NOT included — how to obtain them

The video field in each row is an absolute local path of the form:

file:///YOUR_DATA_ROOT/video-r1-extracted/<DATASET>/<relative_path>

Before using this dataset you must download the source videos and update the paths (or symlink) to match your local layout.

Per-source video download

data_source	Embedded path prefix	Public source	Download
video_r1_llava_mc	.../video-r1-extracted/LLaVA-Video-178K/...	LLaVA-Video-178K (Video-R1 extracted subset)	lmms-lab/Video-R1 — follow extraction instructions in the repo
video_r1_nextqa_mc	.../video-r1-extracted/NeXT-QA/NextQA/NExTVideo/...	NExT-QA	NExT-QA official or Video-R1 hub
video_r1_perceptiontest_mc	.../video-r1-extracted/PerceptionTest/...	PerceptionTest	deepmind/perception_test
video_r1_clevrer_mc	.../video-r1-extracted/CLEVRER/train_videos/...	CLEVRER	CLEVRER project
video_r1_star_mc	.../video-r1-extracted/STAR/...	STAR	STAR benchmark

After downloading, update the video path strings in the parquets (or use a loader that remaps the root prefix) before passing them to a training framework.

---

Reward and answer format

• Reward type: multiple_choice
• Ground truth: single uppercase letter (A, B, C, D, or E)
• Expected model output format:

  <think>...chain-of-thought reasoning...</think>
  <answer>\boxed{B}</answer>
  

• Accuracy rule: extract the letter inside \boxed{...} from the <answer> block; fall back to lmms-eval's extract_mcq_answer priority-ranked extractor if no \boxed{} is present.
• Format reward: 1.0 iff <think>...</think><answer>...</answer> tags are present and the <think> content is ≥ 100 non-whitespace characters; 0.0 otherwise.
• Composite reward used during training: 0.8 × accuracy + 0.2 × format

---

Curation notes

• Derived from lmms-lab/Video-R1 using scripts/data/curate_v1.py --frames 24 --pixels 100000 --out-suffix _24f100k.
• Free-form LLaVA rows (video_r1_llava_free) are excluded — they require an LLM judge and are not in this split.
• Token counts validated to fit within max_prompt_length=4096 for Qwen3-VL.
• Base model for training: Qwen/Qwen3-VL-8B-Instruct (BF16, M-RoPE).

---

Licenses

Source datasets retain their original licenses. Users must comply with each:

Dataset	License
LLaVA-Video / Video-R1	Apache 2.0 (model weights); data follows original YouTube/source terms
NExT-QA	MIT
PerceptionTest	Apache 2.0
CLEVRER	MIT
STAR	Apache 2.0

This repository (parquet structure + curation code) is released under Apache 2.0.

---

Citation

If you use this dataset, please cite the original sources:

@article{videor1,
  title={Video-R1: Reinforcing Video Reasoning in MLLMs},
  author={Feng, Kaituo and others},
  journal={arXiv preprint},
  year={2025}
}

@inproceedings{nextqa,
  title={NExT-QA: Next Phase of Question-Answering to Explain Temporal Actions},
  author={Xiao, Junbin and others},
  booktitle={CVPR},
  year={2021}
}

@inproceedings{perceptiontest,
  title={Perception Test: A Diagnostic Benchmark for Multimodal Video Models},
  author={Patraucean, Viorica and others},
  booktitle={NeurIPS},
  year={2023}
}

@inproceedings{clevrer,
  title={CLEVRER: Collision Events for Video REpresentation and Reasoning},
  author={Yi, Kexin and others},
  booktitle={ICLR},
  year={2020}
}

@inproceedings{star,
  title={STAR: A Benchmark for Situated Reasoning in Real-World Videos},
  author={Wu, Bo and others},
  booktitle={NeurIPS},
  year={2021}
}