This initiative empowers the open-source community to collaboratively build impactful datasets for AI models. It targets researchers and developers seeking high-quality, community-vetted datasets, offering curated resources and tools to facilitate data creation and annotation.

How It Works

The project comprises two main components: community efforts and cookbook efforts. Community efforts involve hands-on projects guided by Hugging Face, such as prompt ranking and image preference annotation, leveraging community participation to create large-scale datasets. Cookbook efforts provide standalone guides and tools for users to independently build domain-specific or preference-based datasets (DPO, ORPO, KTO).

Quick Start & Requirements

• Community Datasets: Available on the Hugging Face Hub (e.g., data-is-better-together/10k_prompts_ranked, data-is-better-together/open-image-preferences-v1-binarized).
• Cookbook Tools: Instructions and guides are provided within individual project READMEs (e.g., cookbook-efforts/domain-specific-datasets/README.md).
• Dependencies: Primarily relies on Hugging Face libraries and tools. Specific requirements vary per project.

Highlighted Details

• Successfully created and released data-is-better-together/10k_prompts_ranked with over 385 contributors.
• Developed multilingual benchmarks (MPEP) by translating high-quality prompts into Dutch, Russian, and Spanish.
• Generated 10K text-to-image preference pairs for evaluating image generation models.
• Cookbook efforts aim to facilitate domain-specific datasets and preference data (DPO, ORPO, KTO) creation.

Maintenance & Community

• A collaboration between Hugging Face, Argilla, and the open-source ML community.
• Community participation is encouraged via Hugging Face Discord and project-specific READMEs.

Licensing & Compatibility

• Datasets are typically released under permissive licenses allowing for broad use. Specific licenses should be checked on the Hugging Face Hub for each dataset.

Limitations & Caveats

• Some community projects are still in progress or have limited language support for translations. Cookbook efforts are designed for standalone use, with varying levels of community guidance.