Yalda Embedding
A Persian dense retrieval model fine-tuned from Qwen3-Embedding-0.6B with 595.8 million parameters, trained through contrastive learning and relational knowledge distillation for Persian retrieval tasks.
Timeline (last 90 days)
Yalda Embedding is a Persian retrieval model with 595.8 million parameters, fine-tuned from Qwen3-Embedding-0.6B and trained on more than one million query-document pairs.
On the eight Persian retrieval tasks of FaMTEB, Yalda Embedding achieves a macro average score of 0.5488, the highest among the 0.6B models evaluated.
Yalda Embedding surpasses Qwen3-Embedding-0.6B and F2LLM-v2-0.6B on every task.
Yalda Embedding approaches the performance of Qwen3-Embedding-4B, winning on four tasks and tying on one, with a macro average only 0.0075 behind.
The final distillation configuration outperforms the non-distilled model on all eight tasks, raising the average score from 0.4813 to 0.5488, without increasing model size at inference time.
Yalda Embedding achieves improvements on hard-negative evaluation tasks without explicitly mining or providing hard negatives for training.
All Facts (6)
Yalda Embedding is a Persian retrieval model with 595.8 million parameters, fine-tuned from Qwen3-Embedding-0.6B and trained on more than one million query-document pairs.
50%UnverifiedOn the eight Persian retrieval tasks of FaMTEB, Yalda Embedding achieves a macro average score of 0.5488, the highest among the 0.6B models evaluated.
50%UnverifiedYalda Embedding surpasses Qwen3-Embedding-0.6B and F2LLM-v2-0.6B on every task.
50%UnverifiedYalda Embedding approaches the performance of Qwen3-Embedding-4B, winning on four tasks and tying on one, with a macro average only 0.0075 behind.
50%UnverifiedThe final distillation configuration outperforms the non-distilled model on all eight tasks, raising the average score from 0.4813 to 0.5488, without increasing model size at inference time.
50%UnverifiedYalda Embedding achieves improvements on hard-negative evaluation tasks without explicitly mining or providing hard negatives for training.
50%