yixinL7/BRIO

Combines contrastive learning with maximum likelihood estimation to train abstractive summarization models, where the model learns to rank system-generated candidate summaries by quality rather than treating all non-reference outputs equally. Built on Hugging Face Transformers with modified BART and PEGASUS implementations for efficient training, and includes preprocessing pipelines for CNN/DailyMail, XSum, and NYT datasets with multi-candidate summary generation. Supports both generation and reranking modes, with standard ROUGE evaluation and checkpoint management for multi-GPU training.

337 stars. No commits in the last 6 months.