multilingual-retrieval-lab-e5s

intfloat/multilingual-e5-small fine-tuned with InfoNCE on GermanDPR (article-disjoint from its test split) and a SQuAD sample, with BM25-mined hard negatives and 50 % machine-translated German queries for English passages (seed 1 of 3). Use the e5 prefixes query: and passage: , mean pooling, cosine similarity.

nDCG@10 (%) GermanDPR test XQuAD en->en XQuAD de->en
multilingual-e5-small 78.0 87.8 76.2
this model (seed 1) 80.5 87.4 78.8

XQuAD is scored against all 2,069 SQuAD v1.1 dev paragraphs. Training data licences: GermanDPR CC BY 4.0, SQuAD CC BY-SA 4.0. Code, all runs and limits: https://github.com/pavanyadava007/multilingual-retrieval-lab (author: Pavan Yadava Annappa).

Downloads last month
26
Safetensors
Model size
0.1B params
Tensor type
F32
·
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Model tree for pavanyadava07/multilingual-retrieval-lab-e5s

Finetuned
(203)
this model

Datasets used to train pavanyadava07/multilingual-retrieval-lab-e5s

Space using pavanyadava07/multilingual-retrieval-lab-e5s 1