Vikhr-YandexGPT-5-Lite-8B-it Overview

Vikhr-YandexGPT-5-Lite-8B-it is an 8 billion parameter instruction-tuned model developed by Vikhrmodels, built upon the YandexGPT-5-Lite-8B-pretrain architecture. It is primarily specialized for the Russian language but offers bilingual RU/EN support.

Key Capabilities & Training

• Instruction Following: The model was fine-tuned using Supervised Fine-Tuning (SFT) on a large (150k instructions) synthetic Russian dataset, GrandMaster-PRO-MAX, which incorporates Chain-Of-Thought (CoT) reasoning.
• RAG Grounding: It includes a dedicated synthetic dataset, Grounded-RAG-RU-v2 (50k dialogues), designed to enhance its ability to provide answers grounded in provided documents. This RAG mode requires a specific system prompt and allows for document content in Markdown, HTML, or Plain Text.
• Bilingual Support: While specialized in Russian, the model supports both Russian and English.

Usage Considerations

• Safety: The model has a low level of response safety, prioritizing instruction adherence. Users should implement safety measures and test thoroughly.
• System Prompts: System prompts are best used for specifying response style (e.g., "answer only in json format") and are recommended to be in English, as this aligns with the training data.
• RAG Mode: Requires a specific GROUNDED_SYSTEM_PROMPT and works best with low temperatures (0.1-0.5) and top_k (30-50) to avoid generation defects.