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}, 
}