SetFit v1.2.0 released: prompt-free few-shot classifier catches up with modern stack

Hugging Face's tomaarsen released SetFit v1.2.0 on September 4 — a prompt-free few-shot text classification framework: it constructs positive/negative pairs from a small number of labeled sentences, contrastively fine-tunes a Sentence Transformer, then trains a classification head on the embeddings, requiring no prompts or LLMs. With just 8 samples per class, it can match the performance of fine-tuning RoBERTa-Large with roughly 3,000 samples.

Confirmed

Why it matters

SetFit lets teams without large amounts of labeled data or those who prefer not to rely on LLMs and prompt engineering train competitive text classifiers with just 8 samples per class. Previous versions had become nearly unusable in the current ecosystem due to aging dependencies; v1.2.0 makes it work again on the latest transformers, huggingfacehub, and Sentence Transformers stacks.

2026-09-04 ~ 2026-09-04 · 9 related posts

Primary sources

1 near-duplicate retellings: tomaarsen