Why did we open-source our inference engine? Read the post

BAAI/bge-reranker-v2-m3

More details please refer to our Github: FlagEmbedding.

Overview

Architecture
XLM-RoBERTa
Parameters
568M
Tasks
Score
Outputs
Score
Max Sequence Length
8,192 tokens
License
apache-2.0
Languages
multilingual

Benchmarks

AskUbuntuDupQuestions

technology reranking en

Duplicate question detection from AskUbuntu

Corpus: 6,743 Queries: 360
Quality
ndcg at 10 0.6763
map at 10 0.5245
mrr at 10 0.7611
Performance L4 b1 c16
Query 6.5K tok/s
Query p50 40.6ms
Reference →

CMedQAv1Reranking

medical reranking zh

Chinese medical question answering reranking (v1)

Corpus: 100,000 Queries: 2,000
Quality
map at 10 0.8192
mrr at 10 0.8550
Reference →

CMedQAv2Reranking

medical reranking zh

Chinese medical question answering reranking (v2)

Corpus: 108,000 Queries: 4,000
Quality
map at 10 0.8287
mrr at 10 0.8608
Reference →

CQADupstackPhysicsRetrieval?candidates_model=Alibaba-NLP

general retrieval en

Performance L4 b1 c16
Query 26.5K tok/s
Query p50 89.8ms

CosQA?candidates_model=Alibaba-NLP

general retrieval en

Performance L4 b1 c16
Query 13.9K tok/s
Query p50 66.1ms

FiQA2018?candidates_model=Alibaba-NLP

general retrieval en

Performance L4 b1 c16
Query 30.0K tok/s
Query p50 90.2ms

LegalBenchConsumerContractsQA?candidates_model=Alibaba-NLP

general retrieval en

Performance L4 b1 c16
Query 51.3K tok/s
Query p50 93.5ms

MMarcoReranking

general reranking zh

Multilingual MARCO passage reranking (Chinese)

Quality
map at 10 0.3354
mrr at 10 0.3401
Performance L4 b1 c16
Reference →

NFCorpus

medical retrieval en

Biomedical literature search from NutritionFacts.org

Corpus: 3,593 Queries: 323
default_bm25-k-100
Quality
ndcg at 10 0.3285
map at 10 0.2339
mrr at 10 0.5301
default_bm25-k-200_qrels-k-50
Quality
ndcg at 10 0.3674
map at 10 0.2602
mrr at 10 0.5765
default_bm25-k-200
Quality
ndcg at 10 0.3280
map at 10 0.2331
mrr at 10 0.5264
default_bm25-k-50_qrels-k-10
Quality
ndcg at 10 0.4029
map at 10 0.2832
mrr at 10 0.6087
default_bm25-k-50
Quality
ndcg at 10 0.3306
map at 10 0.2375
mrr at 10 0.5380
default_bm25-k-100_qrels-k-10
Quality
ndcg at 10 0.3759
map at 10 0.2621
mrr at 10 0.5876
default_bm25-k-50_qrels-k-50
Quality
ndcg at 10 0.4243
map at 10 0.3065
mrr at 10 0.6220
default_bm25-k-200_qrels-k-10
Quality
ndcg at 10 0.3557
map at 10 0.2497
mrr at 10 0.5677
default_bm25-k-100_qrels-k-50
Quality
ndcg at 10 0.3930
map at 10 0.2792
mrr at 10 0.5996
Reference →

NFCorpus?candidates_model=Alibaba-NLP

general retrieval en

Performance L4 b1 c16
Query 35.3K tok/s
Query p50 109.6ms

SCIDOCS?candidates_model=Alibaba-NLP

general retrieval en

Performance L4 b1 c16
Query 29.4K tok/s
Query p50 98.7ms

SciFact?candidates_model=Alibaba-NLP

general retrieval en

Performance L4 b1 c16
Query 33.4K tok/s
Query p50 113.2ms

StackOverflowQA?candidates_model=Alibaba-NLP

general retrieval en

Performance L4 b1 c16
Query 53.1K tok/s
Query p50 111.1ms

T2Reranking

general reranking zh

Chinese passage ranking benchmark

Quality
map at 10 0.5630
mrr at 10 0.7781
Reference →

Self-hosted inference for search & document processing

Cut API costs by 50x, boost quality with 85+ SOTA models, and keep your data in your own cloud.

Github 2.0K

Contact us

Tell us about your use case and we'll get back to you shortly.