---
configs:
- config_name: default
  data_files:
  - split: train_candidate
    path: data/train_candidate-*.parquet
  - split: yt_technical_hidden_v0
    path: data/yt_technical_hidden_v0-*.parquet
---

# youtube_technical_v0

Canonical Phonon ASR dataset. Audio is embedded in Parquet as a Hugging
Face-compatible `audio` struct with `bytes` and `path`.

Hidden eval rows must not be used for training or synthetic prompt generation.


## Hub

- Dataset id: `Infatoshi/youtube_technical_v0` (public)
- Companion labels/manifests/attribution: `Infatoshi/phonon-youtube-technical`
- Rows: see `summary.json` (~118k)
- Audio is embedded in Parquet (`audio` struct: `bytes`, `path`)
- Split `yt_technical_hidden_v0` is a holdout: do not train on it

```bash
hf download Infatoshi/youtube_technical_v0 --repo-type dataset \
  --local-dir /home/infatoshi/phonon/datasets/datasets/youtube_technical_v0
```