SetFit (trained on SBERT) was designed for few-shot learning, but the method can also be applied in scenarios where no (or not enough) labeled data is available for ZERO-Shot classification w/ synthetic data set added. The main trick is to create synthetic examples that resemble the classification task, and then train a SetFit model on them. SetFit can be applied for ZERO-shot classification (with added labeled synthetic data) AND adding synthetic examples can also provide a performance boost to new FEW-shot classification!