Researchers have developed Saidzoda-Tajik-E5-Large, a 560 million parameter text embedder specifically fine-tuned for the Tajik language. This model significantly improves retrieval accuracy on Tajik legal and tax documents, achieving a 93.6% success rate for finding the correct passage among the top ten results for law documents and 94.5% for tax documents. The fine-tuning process involved training on over 235,000 question-passage pairs, incorporating hard negatives and mixed-language data to maintain performance in English and Russian. While the model performs comparably to larger models on specific tasks, its performance on a general test set shows smaller gains, and the researchers note that training in bfloat16 format was unsuccessful. AI
IMPACT Enhances information retrieval capabilities for under-resourced languages like Tajik, potentially enabling new applications in legal and administrative sectors.
RANK_REASON The cluster describes a new research paper detailing the fine-tuning of an existing language model for a specific language and task. [lever_c_demoted from research: ic=1 ai=1.0]
- arXiv
- Belebele
- English
- Giga-Embeddings-3B
- Hugging Face
- multilingual-e5-large-instruct
- Russian
- Saidzoda-Tajik-E5-Large
- Tajik
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →
