---
license: apache-2.0
pretty_name: Test500
---

# Test500

Tar-sharded speech / TTS dataset with a custom Hugging Face loader.
Designed for **streaming**, **low memory usage**, and **raw byte access**.

## Loading

```python
from datasets import load_dataset

ds = load_dataset(
    "thangylvp/test500",
    split="train",
    trust_remote_code=True,
)
```

## Sample format

Each sample corresponds to one `(channel, id)`:

- `channel` (string)
- `id` (string)
- `transcribe_assembly` (string)
- `tokenize_vibevoice` (bytes)

## Notes

- Data is stored in `batch_*.tar` shards
- Files for each sample are written contiguously in tar
- Audio and binary features are returned as raw bytes