Yalda Embedding released as a 595.8M Persian retrieval model
Original title: Yalda-Embedding: Open Development of a Competitive Persian Retrieval Model
Adapted from Qwen3-Embedding-0.6B on more than one million pairs, it scores 0.5488 macro-average on eight FaMTEB tasks.
Authors present Yalda Embedding, a Persian dense retrieval model with 595.8 million parameters, adapted from Qwen3-Embedding-0.6B using more than one million query-document pairs. On eight FaMTEB tasks it posts the highest macro-average among evaluated 0.6B models at 0.5488, beating Qwen3-Embedding-0.6B and F2LLM-v2-0.6B on every task. It trails Qwen3-Embedding-4B by 0.0075 in macro-average while winning four tasks and tying one. Distillation from F2LLM-v2-8B lifts the mean from 0.4813 to 0.5488. Weights, code, and the pairwise dataset are released.
Unconfirmed: FaMTEB scores are the authors' own evaluation, not an independent leaderboard.
- Contrastive fine-tuning alone was modest and regressed on two tasks.
- Distillation improved all eight tasks without a larger inference model.
- Gains cover hard-negative tasks though hard negatives were not mined.
- The release includes a teacher-embedded copy of the dataset.
From the original article
1 min read at arxiv.orgDespite 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.
By Yalda Taheri · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.