Argilla 2.4 release notes / what's new
Argilla 2.4 enables users to prepare AI datasets without writing code by allowing the direct import of datasets from the Hugging Face Hub into Argilla's UI. This update streamlines the collection of human feedback for fine-tuning and evaluation, democratizing dataset creation for domain experts who may not be proficient in Python.
No-Code Dataset Preparation on the Hugging Face Hub
Argilla 2.4 allows users to import any public dataset from the Hugging Face Hub—which currently hosts over 230,000 datasets—to serve as a foundation for AI projects. By using the Argilla UI, developers and domain experts can define questions and collect human feedback without needing to use the Python SDK.
This capability is designed to support several key use cases:
- Community Contribution: Users can import an open dataset into a public Argilla Space and share the URL to gather feedback from the broader community.
- New Dataset Creation: Users can upload a CSV to the Hub and import it into an Argilla Space to begin labeling from scratch.
- Dataset Curation: Existing Hub datasets can be imported into Argilla Spaces for curation specifically for model fine-tuning or evaluation.
- Dataset Improvement: Users can import existing Hub datasets to provide feedback and improve them for the benefit of the community.
Technical Workflow and Implementation
To use these features, users must first deploy Argilla, with the recommended method being deployment on Hugging Face Spaces. The default deployment includes Hugging Face OAuth, which allows any Hub user to contribute annotations if the Space is public.
The import process follows these steps:
- Dataset Selection: Users sign in and use the "Import dataset from Hugging Face" button on the Home page, providing either a repository ID or selecting an example dataset.
- Configuration: Argilla automatically suggests an initial configuration based on the dataset's features. Users can then add questions (such as labels, ratings, rankings, or text) or remove unnecessary fields (such as text, chats, or images).
- Dataset Creation: Once the configuration is finalized, and changes are reflected in real-time, the user clicks "Create dataset" to complete the import.
In the first version of this feature, the Hub dataset must be public. Support for private datasets is currently a requested feature for future development.
Integration with the Hugging Face Ecosystem
While the no-code UI simplifies the import process, the Argilla Python SDK remains available for users who require advanced customization. Argilla is an open-source, data-centric tool integrated with the Hugging Face Hub, designed to facilitate collaboration between AI developers and domain experts to build high-quality datasets.