Spaces:
No application file
No application file
File size: 2,759 Bytes
730f9d2 34bf0cd | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 | ---
title: README
emoji: 🏢
colorFrom: purple
colorTo: yellow
sdk: docker
pinned: false
---
# Reuben Data Lab
> 🏆 Work here was produced for the
> **[Uncharted Data Challenge](https://www.adaptionlabs.ai/blog/the-uncharted-data-challenge)**
> hosted by **[Adaption Labs](https://www.adaptionlabs.ai)** — credit to
> **Adaptive Data by Adaption** for organizing the hackathon.
Building **open, underserved datasets** for training and evaluating modern
audio, speech, and multimodal models. Every release is open-sourced on
Hugging Face with permissive licensing and rich metadata, targeting the three
criteria the Uncharted Data Challenge cares about: **under-served problem
domains**, **scarce open-source data**, and **under-resourced languages**.
## Datasets
### 🎵 [FMA Labeled — Multi-Attribute Music Dataset](https://huggingface.co/datasets/Reubencf/fma-labeled)
29k Creative-Commons tracks from the Free Music Archive, automatically
annotated with **lyrics, genre, sub-genres, mood, instruments, BPM, key,
vocal type, energy, era, and audio quality** using `gemini-flash-latest`.
Paired audio + text for music tagging, music-LM training, and auto-lyric
research.
### 🗣️ [Multilingual Synthetic TTS (Qwen3)](https://huggingface.co/datasets/Reubencf/multilingual-synthetic-tts)
~69k synthetic speech clips across **9 languages** (en, ja, zh, ko, de, es,
fr, ru, pt) generated with Qwen3-TTS-12Hz via zero-shot voice cloning from a
rotating pool of reference speakers. Covers conversational, informational,
technical, emotional, and proverb-style utterances — useful for TTS
fine-tuning, ASR augmentation, and cross-lingual voice-conversion research.
## Focus Areas
- **Under-resourced languages** — expanding speech and text coverage beyond
English-only datasets.
- **Rich supervision** — datasets ship with detailed structured metadata
(genre/mood/BPM/key for music; language/style/voice for speech), not just
audio + class labels.
- **Permissive licensing** — Creative Commons / CC0 where possible; synthetic
outputs released for open research.
- **Reproducibility** — generation pipelines and labeling scripts are
open-sourced alongside the data.
## Tooling & Pipeline
- **Labeling**: Google Gemini (`gemini-flash-latest`) via Flex and Batch APIs.
- **Speech synthesis**: Qwen3-TTS-12Hz-1.7B-Base on 2× H100 with zero-shot
voice cloning.
- **Infra**: Hyperbolic GPU rentals, custom stall-watchers for long-running
multi-GPU jobs, Hugging Face Hub for distribution.
## Get In Touch
- Hugging Face: [@Reubencf](https://huggingface.co/Reubencf)
- Datasets home: [ReubenDataLab](https://huggingface.co/ReubenDataLab)
---
*More datasets coming soon as part of the Uncharted Data Challenge submission.*
|