MS Marco MiniLM-L4 V2

cross-encoder/ms-marco-MiniLM-L4-v2 is a cross-encoder reranking model that scores query-passage pairs for information retrieval. It is trained on the MS MARCO Passage Ranking dataset, which contains over one million real, anonymized queries from Bing search logs and human-generated relevance judgments.

The model is designed to be used in a retrieve-and-rerank pipeline: an initial retrieval step (e.g., with ElasticSearch) returns a candidate set of passages, and this cross-encoder re-scores each query-passage pair to produce a refined ranking. The architecture uses a MiniLM-L4 Transformer (4 layers) with a cross-attention mechanism, enabling higher accuracy than single-vector bi-encoders while maintaining practical throughput.

Performance

The table below reports standard retrieval metrics on the TREC Deep Learning 2019 and MS MARCO Passage Reranking development datasets, along with throughput on a V100 GPU.

The MiniLM-L4-v2 achieves a strong balance of accuracy and speed. On the TREC DL 2019 benchmark, its NDCG@10 of 73.04 places it within the top tier of compact reranking models, while its throughput of 2,500 documents per second on a V100 makes it suitable for latency-sensitive applications.