Summer has wrapped up at Hugging Face, and the season brought a flurry of activity. The Hub expanded from 10,000 to over 16,000 public model repositories, alongside a host of new features, community initiatives, and open-source developments.
New Features
Spaces Beta: Hugging Face launched Spaces, a free platform for hosting machine learning demo apps using Gradio or Streamlit. Users can deploy apps with secrets management, custom dependencies, and GitHub integration. Over 100 Spaces are already live.
Likes: Users can now like models, datasets, or Spaces, showing appreciation for community contributions.
TensorBoard Integration: Automatic TensorBoard instances are now launched for repos containing TensorBoard traces, supporting both public and private repos.
Metrics: Model repos can display evaluation metrics via the model-index section in their model card. These metrics automatically link to Papers With Code leaderboards for easy comparison.
New Widgets: The Hub introduced feature extraction and sentence similarity widgets, bringing the total to 18 interactive widgets for testing models in-browser.
Community
The Hugging Face Course continued to educate developers, and a JAX/FLAX sprint brought contributors together to build and share models.
Open Source
Transformers: New model architectures and improvements were added. Datasets: Enhanced functionality for data loading and processing. Hub: Welcomed new libraries to the ecosystem, expanding integration.
Solutions
Upcoming offerings include Infinity for optimized inference, hardware acceleration options, SageMaker integration, and AutoNLP in the browser. The Inference API and Expert Acceleration Program further support enterprise needs.
Research
Hugging Face continued contributing to cutting-edge AI research, though details were not expanded in this post.
With a focus on open collaboration, Hugging Face’s summer progress underscores its commitment to democratizing machine learning.