ElyKok/hebrew-qa-retriever-e5-large

Hebrew QA Retriever (fine-tuned multilingual-e5-large)

Retrieves the Hebrew paragraph that answers a Hebrew question. Fine-tuned from intfloat/multilingual-e5-large on the answerable questions of HeQ, with 2 mined hard negatives per question and CachedMultipleNegativesRankingLoss (batch 64, 1 epoch, lr 1e-05).

Usage

from sentence_transformers import SentenceTransformer
model = SentenceTransformer("ElyKok/hebrew-qa-retriever-e5-large")
q = model.encode("query: מי ניסח את מגילת העצמאות?", normalize_embeddings=True)
p = model.encode(["passage: ..."], normalize_embeddings=True)

Use the query: and passage: prefixes, as with the base model. For best results, combine it with BM25 through Reciprocal Rank Fusion (the "Hybrid" rows below).

Results (HeQ test set, 1,072 questions, search over all 4,930 paragraphs)

Model	R@1	R@5	R@10	MRR@10	NDCG@10	R@5 Wikipedia	R@5 Geektime
BM25 (keyword search)	73.1	86.8	89.8	79	81.7	91.2	82.2
multilingual-e5-base	67.4	81.7	86.2	73.4	76.5	86.2	77
multilingual-e5-large	72.4	87.8	91.2	79	82	90.8	84.6
Qwen3-Embedding-0.6B	56	71.1	76.9	62.6	66	74.9	67.2
Fine-tuned e5-base (ours)	67.3	82.8	86.5	74.1	77.1	87	78.6
Fine-tuned e5-large (ours)	78.9	92	94.8	84.8	87.3	93.2	90.7
Hybrid: BM25 + e5-large (no training)	80.4	92.1	95.1	85.5	87.8	94.7	89.4
Hybrid: BM25 + fine-tuned e5-large (ours)	82.9	93.2	95.4	87.4	89.4	95.4	90.9

Test articles are disjoint from training articles. HeQ is licensed CC BY 4.0 (Webiks / MAFAT). Limitation: HeQ questions were written while looking at the paragraph, so they share more wording with it than real user queries would.