ngqtrung/mmrl-video-val-v1

    ---
    language:
    - en
    license: apache-2.0
    size_categories:
    - 1K<n<10K
    task_categories:
    - video-text-to-text
    tags:
    - multimodal-rl
    - qwen3-vl
    - gspo
    - grpo
    ---

    # ngqtrung/mmrl-video-val-v1

    Curated v1 dataset for multimodal RL fine-tuning of Qwen3-VL-4B-Instruct.

    | Property | Value |
    |---|---|
    | Rows | 1720 |
    | Modality | video |
    | Split | validation |
    | Schema | verl-ready (prompt + images + videos + reward_model + extra_info) |

    ## Config

    ```json
    {

"fps": 1, "max_frames": 64, "max_pixels": 501760 } ```

    ## System prompt

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

step-by-step reasoning \boxed{final_answer}

Examples:

Q: 7 × 8? 7 × 8 = 56. \boxed{56}

Q: A right triangle has legs of length 3 and 4. What is the hypotenuse? By the Pythagorean theorem, c² = 3² + 4² = 9 + 16 = 25, so c = 5. \boxed{5}

For multiple-choice, put the letter, e.g. \boxed{B}. Always wrap reasoning in ... and answer in \boxed{...}. No text outside these tags. ```

    ## Source credit

    VideoMME v1, VideoMME v2, PerceptionComp, Video-Holmes — eval-only assembly

    ## How to use with verl

    The parquet contains `videos` with `file://` paths pointing to local disk (if video modality). Set `VIDEO_R1_ROOT=/path/to/your/video-r1-extracted` and either edit the parquet's video paths or use a path-resolver wrapper.

    For image rows, images are inline PNG/JPEG bytes — no external dependencies.

    ## Schema

    ```python
    prompt: list[struct(role: str, content: str)]
    images: list[struct(bytes: binary, path: str)]
    videos: list[struct(type, video, fps, max_frames, max_pixels)]
    data_source: str
    ability: str
    reward_model: struct(ground_truth: str, style: str)
    extra_info: struct(reward_type: str, answer: str)
    ```

    ## License

    This dataset inherits the Apache-2.0 license of the underlying sources. See the source credit above for upstream attribution.