SetFit is an efficient, prompt-free framework for few-shot fine-tuning of Sentence Transformers, targeting developers and researchers needing high-accuracy text classification with minimal labeled data. It offers faster training and multilingual capabilities compared to prompt-based methods.

How It Works

SetFit leverages Sentence Transformers to generate rich text embeddings directly, bypassing the need for handcrafted prompts or verbalizers. It employs a two-stage fine-tuning process: first, it trains a classification head on generated embeddings, and then it fine-tunes the entire Sentence Transformer model using these initial predictions. This approach yields competitive accuracy with significantly less data and computation.

Quick Start & Requirements

• Install via pip: pip install setfit
• For bleeding-edge: pip install git+https://github.com/huggingface/setfit.git
• Requires Python 3.9+ (for developer install).
• See Documentation and quickstart for detailed examples.

Highlighted Details

• Achieves high accuracy with as few as 8 labeled examples per class.
• Significantly faster training and inference than large prompt-based models.
• Supports multilingual classification by using any Sentence Transformer checkpoint.
• Integrates seamlessly with the Hugging Face Hub for model sharing and loading.

Maintenance & Community

• Developed by Hugging Face.
• Code formatting enforced by black and isort.
• See notebooks and tutorials for more examples.

Licensing & Compatibility

• Licensed under Apache 2.0.
• Compatible with commercial use and closed-source linking.

Limitations & Caveats

The project is primarily focused on text classification tasks. While it supports multilingual models, performance may vary across languages depending on the underlying Sentence Transformer checkpoint.