Researchers have developed Yalda-Embedding, a new Persian retrieval model with 595.8 million parameters, adapted from Qwen3-Embedding-0.6B. This model demonstrates strong performance on eight Persian retrieval tasks from FaMTEB, achieving the highest macro-average among evaluated 0.6B models and outperforming its base model and F2LLM-v2-0.6B. The development process involved contrastive fine-tuning and relational knowledge distillation, with the final model approaching the performance of a larger Qwen3-Embedding-4B model without an increase in inference-time size. The project emphasizes open development, releasing model weights, code, and datasets to facilitate further research and reuse. AI
IMPACT Enhances retrieval capabilities for Persian language AI applications and provides open-source resources for further research.
RANK_REASON Academic paper detailing a new language model for retrieval tasks. [lever_c_demoted from research: ic=1 ai=1.0]
- F2LLM-v2-0.6B
- F2LLM-v2-8B
- FaMTEB
- Hugging Face
- Mohammad Hassan Heydari
- Persian
- Qwen3-Embedding-0.6B
- Qwen3 Embedding 4B
- Yalda-Embedding
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
