I’ve been working on autofit2, an automated end-to-end pipeline for few-shot multilingual text classification built on SetFit and Sentence Transformers.
It focuses on low-data regimes (often decent results with just a few dozen labeled examples per class), supports 50+ languages, includes emissions tracking, rich automated model cards, entropy-based bias analysis, and packages models for TorchServe, all driven from a single JSON config.
Feedback is welcome.