Yalda-Embedding: Open Development of a Competitive Persian Retrieval Model
arXiv cs.AI · 24 Oct 2026, 03:30 AEST
Yalda Embedding is a new Persian retrieval model that significantly improves performance on various tasks. It adapts from Qwen3-Embedding-0.6B and achieves the highest macro-average among evaluated models.
Key facts
- Yalda Embedding is 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.
- 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.
Read aloud
Yalda Embedding is a new Persian retrieval model that significantly improves performance on various tasks. It adapts from Qwen3-Embedding-zero.6B and achieves the highest macro-average among evaluated models. Yalda Embedding is a five hundred ninety five point eight-million-parameter Persian retrieval model adapted from Qwen3-Embedding-zero.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 zero.6B models, scoring zero point five four eight eight and outperforming Qwen3-Embedding-zero.6B and F2LLM-v2-zero.6B on every task. The final distilled configuration improves all eight tasks over the non-distilled model, raising the mean from zero point four eight one three to zero point five four eight eight without increasing inference-time model size. These gains include hard-negative evaluation tasks despite training without explicitly mined or supplied hard negatives.
Sources in this cluster
- Yalda-Embedding: Open Development of a Competitive Persian Retrieval Model
arXiv cs.AI · 24 Oct 2026, 03:30 AEST