Vikhrmodels/Vikhr-Llama-3.2-1B-instruct-GGUF

💨📱 Vikhr-Llama-3.2-1B-instruct

RU

Инструктивная модель на основе Llama-3.2-1B-Instruct, обученная на русскоязычном датасете GrandMaster-PRO-MAX. В 5 раз эффективнее базовой модели, и идеально подходит для запуска на слабых или мобильных устройствах.

EN

Instructive model based on Llama-3.2-1B-Instruct, trained on the Russian-language dataset GrandMaster-PRO-MAX. It is 5 times more efficient than the base model, making it perfect for deployment on low-power or mobile devices.

• HF model

Рекомендуемая температура для генерации: 0.3 / Recommended generation temperature: 0.3.

Метрики на ru_arena_general / Metrics on ru_arena_general

Model	Score	95% CI	Avg Tokens	Std Tokens	LC Score
kolibri-vikhr-mistral-0427	22.41	+1.6 / -1.6	489.89	566.29	46.04
storm-7b	20.62	+2.0 / -1.6	419.32	190.85	45.78
neural-chat-7b-v3-3	19.04	+2.0 / -1.7	927.21	1211.62	45.56
Vikhrmodels-Vikhr-Llama-3.2-1B-instruct	19.04	+1.3 / -1.6	958.63	1297.33	45.56
gigachat_lite	17.2	+1.4 / -1.4	276.81	329.66	45.29
Vikhrmodels-vikhr-qwen-1.5b-it	13.19	+1.4 / -1.6	2495.38	741.45	44.72
meta-llama-Llama-3.2-1B-Instruct	4.04	+0.8 / -0.6	1240.53	1783.08	43.42

Авторы / Authors

• Sergei Bratchikov, NLP Wanderer, Vikhr Team
• Nikolay Kompanets, LakoMoor, Vikhr Team
• Konstantin Korolev, Vikhr Team
• Aleksandr Nikolich, Vikhr Team

@article{nikolich2024vikhr,
  title={Vikhr: The Family of Open-Source Instruction-Tuned Large Language Models for Russian},
  author={Aleksandr Nikolich and Konstantin Korolev and Sergey Bratchikov and Nikolay Kompanets and Artem Shelmanov},
  journal={arXiv preprint arXiv:2405.13929},
  year={2024},
  url={https://arxiv.org/pdf/2405.13929}
}