SetFit: Efficient Few-Shot Learning Without Prompts

Hugging Face · 3 years ago · 7

SetFit is a framework for few-shot text classification that fine-tunes sentence transformers on small labeled datasets without requiring handcrafted prompts. On the Customer Reviews sentiment dataset, SetFit achieved competitive results with only 8 labeled examples per class, matching performance from fine-tuning RoBERTa Large on 3,000 examples. Training SetFit on an NVIDIA V100 takes 30 seconds at a cost of $0.025, compared to 11 minutes and $0.7 for the larger T-Few 3B model on the same task.