naver/splade

Based on the README, here's a technical summary: Learns sparse query/document representations via BERT's MLM head with regularization, enabling efficient inverted-index retrieval while maintaining explicit lexical matching and interpretability advantages over dense methods. Implements multiple training approaches including hard-negative mining, distillation (MarginMSE, self-distil, ensemble-distil), and query-specific regularization; supports disjoint query/document encoders for efficiency parity with BM25. Built on Hydra for experiment management with end-to-end training, indexing, and retrieval pipelines integrated with MS MARCO and BEIR benchmarks, plus pre-trained models available on Hugging Face.

984 stars. No commits in the last 6 months.