LM-BFF

Why we included this project

LM-BFF is the official code for the ACL 2021 paper on making pre-trained language models work with very few labeled examples. Researchers and practitioners fine-tuning BERT- or RoBERTa-style models on small datasets will find a concrete recipe here: prompt-based fine-tuning with automatically searched templates and demonstrations placed in context. The repo ships scripts to generate k-shot splits from fifteen standard classification datasets, run multi-seed experiments, and aggregate mean and standard deviation results, so it is a practical starting point for reproducing few-shot baselines or building your own prompt experiments. Treat it as a research artifact rather than a maintained library; you will need to adapt the code and pin the older transformers version it targets. It is a solid reference for anyone studying or benchmarking few-shot NLP.