**Vikhr: The Family of Open-Source Instruction-Tuned Large Language Models for Russian** This paper introduces Vikhr, a state-of-the-art open-source instruction-tuned Large Language Model (LLM) specifically designed for the Russian language. Vikhr addresses the challenges of text generation for non-English languages, such as poor quality and computational inefficiency due to tokenization issues. Unlike previous efforts that use LoRA adapters on English-oriented models, Vikhr features an adapted tokenizer vocabulary and undergoes continued pre-training and instruction tuning of all weights. This approach enhances performance and computational efficiency. Vikhr outperforms other open-source and proprietary models on various Russian benchmarks, demonstrating its effectiveness and efficiency. **Contributions:** 1. **Vikhr Model:** A state-of-the-art open-source instruction-following LLM for Russian, with high generation quality and efficient tokenization. 2. **LLM Adaptation Pipeline:** A pipeline for adapting English-oriented LLMs to Russian, including vocabulary adaptation, continued pre-training, and instruction tuning. 3. **Dataset Expansion:** Extensive expansion of Russian datasets for continued pre-training and instruction tuning. 4. **Evaluation:** Vikhr achieves new state-of-the-art results on Russian benchmarks, outperforming other open-source and proprietary models. **Related Work:** The paper reviews existing Russian LLMs, such as ruGPT, Saiga, and ruadapt, highlighting their limitations and the need for more efficient and effective models. **LLM Construction Pipeline:** 1. **Vocabulary Adaptation:** Rebuilds the tokenizer using a language-specific corpus to improve tokenization efficiency. 2. **Continued Pre-training:** Trains the model on large Russian datasets to mitigate vocabulary shift and introduce culture-specific knowledge. 3. **Instruction Tuning:** Extends the Saiga dataset with automatically translated and cleaned English instruction datasets, improving zero-shot performance. **Experiments:** Vikhr is evaluated on multiple benchmarks, including MMLU, Ru-MMLU, CheGeKa, Russian SuperGLUE, and MERA. It outperforms all open-source models and some proprietary closed-source models, demonstrating its superior performance and computational efficiency. **Conclusion:** Vikhr is a comprehensive solution for Russian LLMs, offering high-quality text generation and efficient computational performance. The model's availability and open-source nature aim to foster further research and enhance the diversity of languages in LLMs.**Vikhr: The Family of Open-Source Instruction-Tuned Large Language Models for Russian** This paper introduces Vikhr, a state-of-the-art open-source instruction-tuned Large Language Model (LLM) specifically designed for the Russian language. Vikhr addresses the challenges of text generation for non-English languages, such as poor quality and computational inefficiency due to tokenization issues. Unlike previous efforts that use LoRA adapters on English-oriented models, Vikhr features an adapted tokenizer vocabulary and undergoes continued pre-training and instruction tuning of all weights. This approach enhances performance and computational efficiency. Vikhr outperforms other open-source and proprietary models on various Russian benchmarks, demonstrating its effectiveness and efficiency. **Contributions:** 1. **Vikhr Model:** A state-of-the-art open-source instruction-following LLM for Russian, with high generation quality and efficient tokenization. 2. **LLM Adaptation Pipeline:** A pipeline for adapting English-oriented LLMs to Russian, including vocabulary adaptation, continued pre-training, and instruction tuning. 3. **Dataset Expansion:** Extensive expansion of Russian datasets for continued pre-training and instruction tuning. 4. **Evaluation:** Vikhr achieves new state-of-the-art results on Russian benchmarks, outperforming other open-source and proprietary models. **Related Work:** The paper reviews existing Russian LLMs, such as ruGPT, Saiga, and ruadapt, highlighting their limitations and the need for more efficient and effective models. **LLM Construction Pipeline:** 1. **Vocabulary Adaptation:** Rebuilds the tokenizer using a language-specific corpus to improve tokenization efficiency. 2. **Continued Pre-training:** Trains the model on large Russian datasets to mitigate vocabulary shift and introduce culture-specific knowledge. 3. **Instruction Tuning:** Extends the Saiga dataset with automatically translated and cleaned English instruction datasets, improving zero-shot performance. **Experiments:** Vikhr is evaluated on multiple benchmarks, including MMLU, Ru-MMLU, CheGeKa, Russian SuperGLUE, and MERA. It outperforms all open-source models and some proprietary closed-source models, demonstrating its superior performance and computational efficiency. **Conclusion:** Vikhr is a comprehensive solution for Russian LLMs, offering high-quality text generation and efficient computational performance. The model's availability and open-source nature aim to foster further research and enhance the diversity of languages in LLMs.