Persian text search improves with new open embedding model
Yalda-Embedding: Open Development of a Competitive Persian Retrieval Model
Computation and LanguageArtificial Intelligence
Summary
Finding good matches in Persian text has been hard because there aren’t many tools and data for training models. The authors developed Yalda Embedding, a model specifically adapted for Persian that looks at large amounts of question-answer pairs to learn better connections. It works better than similar-sized models on multiple tests and nearly matches much bigger models. The work also shares all the data and code openly for others to check and build on.
What this means in practice
- •For search engine developers: Improve Persian text search quality by integrating Yalda Embedding for more accurate retrieval across various queries without increasing model size at runtime.
- •For natural language processing engineers: Use the released data and code to build or enhance Persian language models that better understand user queries and documents via dense retrieval techniques.
Authors
Yalda Taheri, Mohammad Hassan Heydari, Armon Rasooli
Abstract
Despite advances in multilingual text representations, effective dense retrieval for Persian remains constrained by the limited availability of language-specific training resources and compact embedding models. We present Yalda Embedding, a 595.8-million-parameter Persian retrieval model adapted from Qwen3-Embedding-0.6B using more than one million query-document pairs. On eight Persian retrieval tasks from FaMTEB, the model achieves the highest macro-average among the evaluated 0.6B models, scoring 0.5488 and outperforming Qwen3-Embedding-0.6B and F2LLM-v2-0.6B on every task. It also approaches Qwen3-Embedding-4B, winning on four tasks, tying on one, and trailing by 0.0075 in macro-average. An incremental ablation study examines contrastive batch size, optimization stabilization, and relational knowledge distillation from F2LLM-v2-8B. Under the one-epoch training setup, contrastive fine-tuning alone yields a modest mean improvement over initialization and regressions on two tasks. The final distilled configuration improves all eight tasks over the non-distilled model, raising the mean from 0.4813 to 0.5488 without increasing inference-time model size. These gains include hard-negative evaluation tasks despite training without explicitly mined or supplied hard negatives. We release model weights, training and evaluation code, the pairwise dataset, and its teacher-embedded counterpart, enabling inspection of data composition and training procedures, independent evaluation, and reuse without repeating expensive teacher inference.