README.md

metadata

configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
dataset_info:
  features:
    - name: chosen
      dtype:
        audio:
          sampling_rate: 44100
    - name: reject
      dtype:
        audio:
          sampling_rate: 44100
    - name: captions
      dtype: string
    - name: duration
      dtype: int32
    - name: iteration
      dtype: int32
  splits:
    - name: train
      num_bytes: 180239660645
      num_examples: 100000
  download_size: 172620977911
  dataset_size: 180239660645
task_categories:
  - text-to-audio
tags:
  - DPO
  - text-to-audio

Dataset Description

This dataset consists of 100k audio preference pairs generated by TangoFlux during the CRPO stage. Specifically, TangoFlux performed five iterations of CRPO. In each iteration, 20k prompts were sampled from a prompt bank. For each prompt, audio samples with the highest and lowest CLAP scores were selected to form the "chosen" and "rejected" pairs, respectively. This process resulted in a total of 100k preference pairs.

Since every iteration contains 20k prompts sampled from audiocaps prompts, some prompts are the same across iterations.

Dataset Sources

• Repository: https://github.com/declare-lab/TangoFlux
• Paper : https://arxiv.org/abs/2412.21037
• Demo : https://huggingface.co/spaces/declare-lab/TangoFlux

Uses

You can directly download the dataset and use them for preference optimization in text-to-audio.

Citation

If you find our dataset useful, please cite us! Thanks!

BibTeX:

@misc{hung2024tangofluxsuperfastfaithful,
      title={TangoFlux: Super Fast and Faithful Text to Audio Generation with Flow Matching and Clap-Ranked Preference Optimization}, 
      author={Chia-Yu Hung and Navonil Majumder and Zhifeng Kong and Ambuj Mehrish and Rafael Valle and Bryan Catanzaro and Soujanya Poria},
      year={2024},
      eprint={2412.21037},
      archivePrefix={arXiv},
      primaryClass={cs.SD},
      url={https://arxiv.org/abs/2412.21037}, 
}