CRPO
收藏资源简介:
CRPO数据集是由新加坡科技设计大学和NVIDIA的研究团队通过CLAP-Ranked Preference Optimization (CRPO)框架创建的音频偏好数据集。该数据集旨在解决文本到音频生成模型在偏好对齐方面的挑战,通过迭代生成和优化偏好数据来提升模型性能。数据集的内容包括通过CLAP模型排名的音频样本,用于构建偏好对并进行优化。CRPO数据集的应用领域主要集中在文本到音频生成模型的训练和优化,旨在提高生成音频的质量和与文本描述的匹配度。
The CRPO dataset is an audio preference dataset developed by research teams from the Singapore University of Technology and Design and NVIDIA via the CLAP-Ranked Preference Optimization (CRPO) framework. This dataset targets the challenges of preference alignment in text-to-audio generation models, and improves model performance through iteratively generating and optimizing preference data. The dataset includes audio samples ranked by the CLAP model, which are used to build preference pairs for model optimization. The primary application scenarios of the CRPO dataset are the training and optimization of text-to-audio generation models, aiming to enhance the quality of generated audio and its alignment with corresponding textual descriptions.
- 1TangoFlux: Super Fast and Faithful Text to Audio Generation with Flow Matching and Clap-Ranked Preference Optimization新加坡科技设计大学, NVIDIA · 2024年



