| Property | Value |
|---|---|
| Base Model | YandexGPT-5-Lite-8B-pretrain |
| Parameter Count | 8 Billion |
| Training Data | GrandMaster-PRO-MAX, Grounded-RAG-RU-v2 |
| Language Support | Russian and English (Bilingual) |
| Paper | Vikhr: Advancing Open-Source Bilingual Instruction-Following Large Language Models |
What is Vikhr-YandexGPT-5-Lite-8B-it?
Vikhr-YandexGPT-5-Lite-8B-it is an instruction-tuned language model designed specifically for bilingual Russian and English language processing. Built upon YandexGPT-5-Lite-8B-pretrain, this model has been fine-tuned using Supervised Fine-Tuning (SFT) on specialized datasets including GrandMaster-PRO-MAX and Grounded-RAG-RU-v2.
Implementation Details
The model underwent extensive training using two key datasets: a 150k instruction dataset (GrandMaster-PRO-MAX) featuring built-in Chain-of-Thought (CoT) reasoning, and a 50k dialogue dataset (Grounded-RAG-RU-v2) specifically designed for RAG capabilities. The model supports various quantization options including GGUF, MLX, 4-bit, and 8-bit variants.
- Specialized RAG implementation with document-based context handling
- Temperature-sensitive performance (optimal at 0.1-0.5)
- Support for HTML, Markdown, and Plain Text document formats
- Maximum context length handling up to 4k symbols per document
Core Capabilities
- Bilingual instruction following in Russian and English
- Advanced RAG (Retrieval-Augmented Generation) capabilities
- Chain-of-Thought reasoning
- Document grounding and contextual response generation
- Various deployment options through different quantization methods
Frequently Asked Questions
Q: What makes this model unique?
The model's distinctive feature is its specialized bilingual capabilities combined with advanced RAG implementation and Chain-of-Thought reasoning, particularly optimized for Russian and English language processing.
Q: What are the recommended use cases?
The model is particularly well-suited for document-grounded question answering, bilingual instruction following, and applications requiring context-aware responses. It's optimized for deployment scenarios requiring both Russian and English language processing capabilities.