Hugging Face Overview
Overview
Vikhr-Llama-3.2-1B-Instruct is a compact, instruction-tuned language model developed by Vikhrmodels, building upon the Llama-3.2-1B-Instruct architecture. It is specifically fine-tuned for the Russian language using the proprietary GrandMaster-PRO-MAX dataset, which consists of 150k instructions with Chain-Of-Thought (CoT) support generated via GPT-4-turbo prompts.
Key Capabilities & Features
- Russian Language Specialization: Optimized for high performance in Russian language tasks.
- Efficiency: Demonstrates 5 times greater efficiency than its base model, making it suitable for resource-constrained environments.
- Compact Size: With a model size under 3GB, it is designed for deployment on low-power and mobile devices.
- Supervised Fine-Tuning (SFT): Trained using SFT on a synthetic dataset to enhance instruction following.
- Competitive Performance: Achieves a score of 19.04 on the
ru_arena_generalbenchmark, outperforming its base model (4.04) significantly.
Use Cases
This model is ideal for applications requiring a highly efficient and compact Russian-language LLM. Its small footprint and optimized performance make it particularly well-suited for:
- Mobile Applications: Integrating advanced language capabilities into mobile devices.
- Edge Computing: Deploying LLM functionality on devices with limited computational resources.
- Russian-centric NLP Tasks: Instruction-following, text generation, and conversational AI in Russian.