Rerankers benchmarks
Every benchmark the Rerankers index tracks, with each core benchmark's share of the composite and how many models it covers.
21 benchmarks tracked.
- Agentset reranker ELO (GPT-5 pairwise judge, FiQA/SciFact/PG over FAISS top-50)
- BEIR average nDCG@10 (reranking a fixed first-stage top-100)
- BRIGHT average nDCG@10 (12 reasoning-intensive datasets)
- CMTEB-R retrieval subsets (Chinese), top-100 from Qwen3-Embedding-0.6B
- CoIR code retrieval nDCG@10 (reranking)
- FollowIR p-MRR (instruction following)
- Hindsight reranker MRR (LoCoMo conv-43, 165 questions, 300 candidates)
- JinaVDR visual document retrieval (reranking)
- MAIR (heterogeneous instructed retrieval) nDCG@10
- MIRACL (18 languages) nDCG@10 after reranking
- MLDR long-document multilingual reranking
- MMEB-v2 retrieval average (image, video, visual-document)
- MMTEB-R multilingual retrieval subsets, top-100 from Qwen3-Embedding-0.6B
- MMTEB reranking average nDCG@10 (MIRACL, XGLUE, HUME, MLDR, Wikipedia reranking)
- MTEB(Code) retrieval subsets, top-100 from Qwen3-Embedding-0.6B
- MTEB(eng, v2) Retrieval nDCG@10, top-100 rerank, mean over six first-stage embedders
- MTEB(eng,v2) retrieval subsets, reranking top-100 from Qwen3-Embedding-0.6B (MTEB-R)
- MTEB reranking-task main score (per-task, mostly MAP@1000)
- RTEB retrieval (as used in jina-reranker-v3.5 card)
- ViDoRe v1 nDCG@5 (visual document retrieval)
- ViDoRe v3 visual document retrieval (reranking)