README.md

---
license: cc-by-4.0
task_categories:
  - question-answering
language:
  - en
tags:
  - lance
  - video
  - benchmark
formats:
  - lance
libraries:
  - datasets
  - lance
size_categories:
  - 1K<n<10K
---

# MINERVA for lmms-eval

This dataset package is prepared for lmms-eval integration of MINERVA.

## Contents

- `minerva.json`: original annotation metadata.
- `data/train.lance`: Lance table with one row per `video_id` and inline `video_blob`.

## Lance schema

- `video_id`: YouTube video ID.
- `youtube_url`: YouTube URL reconstructed from ID.
- `video_ext`: local downloaded file extension.
- `video_size_bytes`: raw file size.
- `video_blob`: large binary blob with metadata `lance-encoding:blob=true`.

## Loading examples

```python
import lance

ds = lance.dataset("hf://datasets/lmms-lab-eval/minerva/data/train.lance")
row = ds.scanner(columns=["video_id"], limit=1).to_table().to_pylist()[0]
blob = ds.take_blobs("video_blob", ids=[0])[0]
with open("sample.mp4", "wb") as f:
    f.write(blob.read())
```