opensearch-project/opensearch-neural-sparse-encoding-doc-v2-distill
The model should be selected considering search relevance, model inference and retrieval efficiency(FLOPS). We benchmark models' zero-shot performance on a subset of BEIR benchmark: TrecCovid,NFCorpus,NQ,HotpotQA,FiQA,ArguAna,Touche,DBPedia,SCIDOCS,FEVER,Climate FEVER,SciFact,Quora.
Benchmarks
CQADupstackPhysicsRetrieval
Duplicate question retrieval from StackExchange Physics
Corpus: 38,314 Queries: 1,039
Performance L4 b1 c16
Corpus 31.0K tok/s
Corpus p50 60.2ms
Query 4.1K tok/s
Query p50 41.5ms
CosQA
Code search with natural language queries
Corpus: 6,267 Queries: 500
Performance L4 b1 c16
Corpus 17.8K tok/s
Corpus p50 46.4ms
Query 2.6K tok/s
Query p50 37.3ms
FiQA2018
Financial opinion mining and question answering
Corpus: 57,599 Queries: 648
Performance L4 b1 c16
Corpus 43.5K tok/s
Corpus p50 56.2ms
Query 4.2K tok/s
Query p50 42.8ms
LegalBenchConsumerContractsQA
Question answering on consumer contracts
Corpus: 153 Queries: 396
Performance L4 b1 c16
Corpus 104.9K tok/s
Corpus p50 72.6ms
Query 6.9K tok/s
Query p50 37.3ms
NFCorpus
Biomedical literature search from NutritionFacts.org
Corpus: 3,593 Queries: 323
Quality
ndcg at 10 0.3398
map at 10 0.1303
mrr at 10 0.5489
Performance L4 b1 c16
Corpus 56.8K tok/s
Corpus p50 70.8ms
Query 1.8K tok/s
Query p50 41.7ms
SCIDOCS
Citation prediction, document classification, and recommendation for scientific papers
Corpus: 25,656 Queries: 1,000
Performance L4 b1 c16
Corpus 48.7K tok/s
Corpus p50 58.5ms
Query 4.7K tok/s
Query p50 37.7ms
SciFact
Scientific claim verification using research literature
Corpus: 5,183 Queries: 300
Performance L4 b1 c16
Corpus 60.2K tok/s
Corpus p50 66.4ms
Query 5.8K tok/s
Query p50 44.1ms
StackOverflowQA
Programming question answering from Stack Overflow
Corpus: 19,931 Queries: 1,994
Performance L4 b1 c16
Corpus 49.6K tok/s
Corpus p50 70.8ms
Query 77.6K tok/s
Query p50 52.3ms