Saidzoda-CA-EmbeddingGemma-300M
A multilingual embedding model fine-tuned from EmbeddingGemma-300M. A single model simultaneously supports semantic retrieval in four Central Asian languages: Tajik, Kazakh, Uzbek, and Kyrgyz.
Timeline (last 90 days)
Saidzoda-CA-EmbeddingGemma-300M is a single multilingual embedding model fine-tuned from EmbeddingGemma-300M, serving four languages simultaneously: Tajik, Kazakh, Uzbek and Kyrgyz.
The fine-tuned model's training set contains 279,904 question-passage pairs, in which the Kazakh, Uzbek and Kyrgyz questions were synthetically generated by the authors' language model from source passages, mixed with Tajik review data and English and Russian replay data.
On the Belebele dataset, Saidzoda-CA-EmbeddingGemma-300M improved recall@1 for the four languages by 14.4 to 26.0 percentage points, reaching 75.2% to 77.7%.
Compared with monolingual models trained separately under the same settings, the unified model's recall@1 gap on Kazakh, Uzbek and Tajik is no more than 2.8 percentage points, and the confidence intervals for recall@10 overlap.
On the Tajik legal retrieval task, the unified model achieved a recall@10 of 92.2, essentially on par with the starting model's 92.4, but recall@1 dropped by 4.4 percentage points.
The unified model's recall@1 on Russian and English dropped by 3.0 and 4.2 percentage points respectively, a common cost of multilingual fine-tuning.
All Facts (6)
Saidzoda-CA-EmbeddingGemma-300M is a single multilingual embedding model fine-tuned from EmbeddingGemma-300M, serving four languages simultaneously: Tajik, Kazakh, Uzbek and Kyrgyz.
50%UnverifiedThe fine-tuned model's training set contains 279,904 question-passage pairs, in which the Kazakh, Uzbek and Kyrgyz questions were synthetically generated by the authors' language model from source passages, mixed with Tajik review data and English and Russian replay data.
50%UnverifiedOn the Belebele dataset, Saidzoda-CA-EmbeddingGemma-300M improved recall@1 for the four languages by 14.4 to 26.0 percentage points, reaching 75.2% to 77.7%.
50%UnverifiedCompared with monolingual models trained separately under the same settings, the unified model's recall@1 gap on Kazakh, Uzbek and Tajik is no more than 2.8 percentage points, and the confidence intervals for recall@10 overlap.
50%UnverifiedOn the Tajik legal retrieval task, the unified model achieved a recall@10 of 92.2, essentially on par with the starting model's 92.4, but recall@1 dropped by 4.4 percentage points.
50%UnverifiedThe unified model's recall@1 on Russian and English dropped by 3.0 and 4.2 percentage points respectively, a common cost of multilingual fine-tuning.
50%