# Russian optimization

This project uses the mergekit and DARE TIES methods to merge the soob3123/amoral-gemma3-12B-v2 and IlyaGusev/saiga_gemma3_12b models, aiming to provide more powerful language processing capabilities.

This project adapts the T-lite-it-1.0 model to Russian. By replacing the tokenizer, continuing pre-training on the Russian corpus, and applying LEP technology, the generation speed of Russian text is improved.

This project is the result of adapting the model to Russian. By replacing the tokenizer, continuing pre-training on the Russian corpus, and applying LEP technology, the generation speed of Russian text has been improved, providing a new solution for Russian natural language processing.

A Russian instruction-optimized model fine-tuned based on Qwen2.5-1.5B-Instruct, with 1.5B parameters, supporting both Russian and English, excelling in text QA within RAG workflows

A Russian-optimized instruction model based on Llama-3.2-1B-Instruct, with a 5-fold increase in efficiency, suitable for deployment on low-power devices

Cyclone 0.5* is an open-source large language model supporting Russian and English, extending the RoPE context window to 32k and optimizing JSON processing and multi-turn dialogue capabilities

Vikhr is a Russian language model based on the Mistral architecture, undergoing three rounds of pre-training with 400 million tokens. It outperforms the original Mistral in Russian tasks but may have shortcomings in code processing.

A streamlined version based on google/mt5-base, optimized for Russian and English with 58% fewer parameters