---
pretty_name: B150 Public Eval Smoke 100
task_categories:
- text-to-speech
language:
- en
- zh
tags:
- tts
- speech
- controllable-speech-synthesis
- fine-tuning
license: other
---

# B150 Public Eval Smoke 100

This dataset is a 100-sample public-eval smoke split released as raw
audio plus a lightweight JSONL manifest.

It is intended for low-cost pipeline bring-up and local fine-tuning
smoke tests. It is not the recommended split for formal experiments.

## What Is Included

- `raw/audio/...`: the released MP3 audio files
- `selected_samples.exported.jsonl`: sample manifest with
  `audio_path`, `target_text`, `duration`, `language`, and
  `prompt_boundary_index`
- `export_info.json`: lightweight release metadata

This repository no longer ships prebuilt Arrow artifacts. Prepare the
local fine-tuning dataset on your own machine with the corresponding
MAGIC-TTS beta data-preparation scripts.

## Associated Paper

This dataset is part of the `MAGIC-TTS` release:

- Paper: https://arxiv.org/abs/2604.21164
- Project repo: https://huggingface.co/maimai11/MAGIC-TTS

If you use this dataset, please cite the paper above so the dataset page is
linked with the corresponding research release.