Meno-Lite-0.1: A 7B Model for Context-Grounded Tasks

Meno-Lite-0.1, developed by Ivan Bondarenko at Novosibirsk State University, is a 7 billion parameter language model built on the Qwen2.5 architecture. Its core philosophy is to prioritize language skills (comprehension, extraction, reasoning) over factual memorization, making it highly effective for Retrieval-Augmented Generation (RAG) scenarios. This model is specifically designed to excel when provided with external context, rather than relying on its parametric memory for world knowledge.

Key Capabilities

• Superior RAG Performance: Achieves top results among 7B models on MultiQ (multi-hop question answering) and competitive multi-hop reasoning at shorter contexts.
• Knowledge Graph Construction: Ranks #1 on NEREL-bench for knowledge graph construction (named entity recognition, relation extraction, contextual definition generation), outperforming models up to 32B.
• Long-Context Understanding: Demonstrates near-perfect passkey retrieval up to 128k tokens, indicating strong ability to locate information in very long documents.
• Efficient Russian Processing: Features a highly optimized tokenizer for Russian, achieving 3.77 characters per token, a 47% improvement over the original Qwen2.5 tokenizer, leading to faster inference and lower costs for Russian workloads.
• Bilingual Support: Primarily focused on Russian, but retains strong English performance due to its training lineage and data.
• Resource-Efficient Deployment: Designed to fit and operate efficiently on a single consumer GPU.

Good For

• RAG pipelines: Document question answering, retrieval-augmented generation.
• Information extraction: Named entity recognition, relation extraction, structured data extraction.
• Knowledge graph construction: Automated knowledge base building, GraphRAG.
• Document processing: Summarization, analysis of legal or technical documents.
• Applications requiring context-grounded reasoning rather than broad world knowledge.