Hamsan (همسان) — Persian sentence embeddings

Hamsan is a Persian text-embedding model built on intfloat/multilingual-e5-base. It turns Persian sentences and passages into 768-dimensional vectors for semantic search, RAG, clustering, and similarity.

همسان یک مدل embedding فارسی است برای جست‌وجوی معنایی، RAG و سنجش شباهت جمله‌ها.

• Same size and speed as e5-base: 278M parameters, runs on a laptop CPU or a small GPU.
• Open license: CC BY-SA 4.0. No non-commercial training data.
• Short vectors: trained with Matryoshka loss, so vectors can be cut to 512/256/128 dims.
• Built-in prompts: query / passage prefixes are saved in the model.

Usage

from sentence_transformers import SentenceTransformer

model = SentenceTransformer("hanyeh00/hamsan")

queries = model.encode(["پایتخت ایران کجاست؟"], prompt_name="query")
docs = model.encode([
    "تهران پایتخت و پرجمعیت‌ترین شهر ایران است.",
    "اصفهان یکی از شهرهای تاریخی ایران است.",
], prompt_name="passage")

print(model.similarity(queries, docs))

• Questions and search queries: prompt_name="query".
• Documents and passages to search over: prompt_name="passage".
• Comparing two sentences with each other: use "query" for both.

Shorter vectors (smaller index, faster search):

model = SentenceTransformer("hanyeh00/hamsan", truncate_dim=256)

Evaluation

Lite FaMTEB subset (8 retrieval + 2 STS tasks). Retrieval: nDCG@10. STS: Spearman.

Task	e5-base	Hamsan
SynPerChatbotRAGFAQRetrieval	0.285	0.287
WikipediaRetrievalMultilingual	0.881	0.861
MSMARCO-FaHardNegatives	0.667	0.628
ArguAna-Fa.v2	0.338	0.494
FiQA2018-Fa.v2	0.230	0.227
SCIDOCS-Fa.v2	0.118	0.141
SciFact-Fa.v2	0.582	0.582
NeuCLIR2023RetrievalHardNegatives	0.530	0.561
Retrieval average	0.454	0.473
Farsick (STS)	0.686	0.707
SynPerSTS (STS)	0.869	0.853
STS average	0.778	0.780

Hamsan improves the retrieval average over e5-base by about 2 points, with the largest gain on argument retrieval (ArguAna, +0.16) and gains on scientific (SCIDOCS) and news (NeuCLIR) retrieval. It is slightly weaker on web search (MSMARCO) and Wikipedia retrieval, and on par for sentence similarity. Larger models such as BGE-M3 or Hakim score higher on FaMTEB.

Training

1. Fine-tuning e5-base for 1 epoch on ~115k Persian pairs and triplets from several domains:

   • Persian Wikipedia (title → paragraph), PQuAD and PersianQA (question → paragraph)
   • pn_summary news (title / summary → article)
   • persian-med-qa (medical question → answer)
   • synthetic STS triplets

   Hard negatives were mined with the base model. Every batch comes from a single source, so in-batch negatives are from the same domain. Loss: MatryoshkaLoss(CachedMultipleNegativesRankingLoss), batch 128, lr 7e-6, max length 256, token embeddings frozen.

2. Merging with the base model to keep its general knowledge: weights = 0.7 × fine-tuned + 0.3 × e5-base.

Limitations

• Only Persian was evaluated; other languages may be weaker than in e5-base.
• Not a source of medical advice, even though medical Q&A was part of the training data.
• Inputs longer than 256 tokens are truncated.

Citation

If you use this model, please cite the base model:

@article{wang2024multilingual,
  title={Multilingual E5 Text Embeddings: A Technical Report},
  author={Wang, Liang and Yang, Nan and Huang, Xiaolong and Yang, Linjun and Majumder, Rangan and Wei, Furu},
  journal={arXiv preprint arXiv:2402.05672},
  year={2024}
}