README.md

---
task_categories:
  - video-text-to-text
  - visual-question-answering
language:
  - en
pretty_name: Video-MME Base64 (480p H.264)
size_categories:
  - 1K<n<10K
source_datasets:
  - lmms-lab/Video-MME
---

# Video-MME Base64 (480p H.264)

Base64-encoded video dataset derived from [lmms-lab/Video-MME](https://huggingface.co/datasets/lmms-lab/Video-MME).

All videos re-encoded to 480p H.264 for VLM compatibility.

## Structure

| Split | Key | Description |
|-------|-----|-------------|
| `qa/` | `video_id` | QA pairs from Video-MME |
| `videos/` | `video_id` | Base64 video (H.264) |
| `audio/` | `video_id` | Base64 audio (MP3) |

Join on `video_id` (e.g., "001", "002").

## Stats

- Videos: 869
- QA pairs: 2607
- Shards: shard-01-of-10 through shard-10-of-10

## Usage

```python
from datasets import load_dataset
import base64

qa = load_dataset("MrVolts/Video-MME-Base64", data_files="qa/*.parquet", split="train")
videos = load_dataset("MrVolts/Video-MME-Base64", data_files="videos/*.parquet", split="train")

# Join on video_id
sample = qa[0]
video_row = videos.filter(lambda x: x["video_id"] == sample["video_id"])[0]
video_bytes = base64.b64decode(video_row["video_b64"])
```