docs/dataset_adapters.md

# Dataset Adapter Notes

The artifact supports small local subsets. It does not bundle restricted dataset media.

## YouCook2
- Expected media: `data/raw/youcook2/videos/*.mp4`
- Optional annotations: `data/raw/youcook2/annotations.json` or `annotations.csv`
- Useful labels: cooking action, food object, chopping/cutting, recipe step captions.

## AVE
- Expected media: `data/raw/ave/videos/*.mp4`
- Optional annotations: `data/raw/ave/annotations.json` or `annotations.csv`
- Useful labels: visible event and aligned event sound.

## AVA ActiveSpeaker
- Expected media: `data/raw/ava_active_speaker/videos/*.mp4`
- Optional annotations: `data/raw/ava_active_speaker/labels.json` or `labels.csv`
- Useful labels: active speaker, visible face/person, speech.

## TVQA
- Expected media: `data/raw/tvqa/videos/*.mp4`
- Optional subtitles: `data/raw/tvqa/subtitles/*.srt` or JSON annotations.
- Useful labels: subtitle speech and visible speaking/action context.

## ActivityNet Captions
- Expected media: `data/raw/activitynet_captions/videos/*.mp4`
- Optional annotations: `data/raw/activitynet_captions/captions.json`
- Useful labels: captioned actions and scene/event windows.

When these files are absent, `make smoke` and `make run-small` use the generated fixture in `data/fixtures/`. The fixture is documented as test-only in each clip metadata record.