ms-marco-MiniLM-L2-v2

The cross-encoder/ms-marco-MiniLM-L2-v2 model is a compact cross-encoder fine-tuned for passage reranking. Given a query and a set of candidate passages (for example, retrieved via a first-stage retrieval system like ElasticSearch), the model assigns a relevance score to each query–passage pair. Passages can then be sorted by score to produce a reranked result list. This model was trained on the MS Marco Passage Ranking task.

Key Strengths

• Efficient: With only 2 layers (MiniLM-L2 architecture), it offers a favorable speed–quality trade-off, processing approximately 4,100 query–passage pairs per second on a V100 GPU.
• Competitive performance: Achieves an NDCG@10 of 71.01 on the TREC Deep Learning 2019 benchmark and an MRR@10 of 34.85 on the MS Marco Dev set.
• Suitable for production reranking pipelines where low latency and throughput are important.

Performance Benchmark

The table below compares this model with other pre-trained cross-encoders on standard reranking metrics (NDCG@10 and MRR@10) and throughput (docs/second on a V100 GPU).