weizhepei/InstructRAG

Employs instruction-tuned LMs to generate self-supervised rationales that explicitly denoise retrieved passages, improving verifiability without external labels. Supports both in-context learning and supervised fine-tuning workflows across multiple QA benchmarks (PopQA, TriviaQA, Natural Questions, ASQA, 2WikiMultiHopQA) with pluggable retrievers like Contriever, DPR, and BM25. Pre-trained checkpoints and training scripts target multi-GPU setups, enabling 8%+ performance gains and robustness to noisy retrieval contexts.

138 stars. No commits in the last 6 months.