RigoChat-7b-v2: Spanish-Optimized LLM

RigoChat-7b-v2 is the second iteration of a Large Language Model (LLM) family developed by Instituto de Ingeniería del Conocimiento (IIC), built upon the Qwen-2.5-7B-Instruct architecture. This 7.6 billion parameter model is specifically fine-tuned using Direct Preference Optimization (DPO) to deliver superior performance for Spanish language queries and tasks. A key focus of its training was to improve responses in RAG (Retriever Augmented Generation) systems with Spanish databases, aiming to prevent hallucinations and ensure safer, more accurate outputs.

Key Capabilities & Features

• Spanish Language Specialization: Optimized for a wide range of NLP tasks with Spanish instructions, including Tool Use, Summarization, Math, Code, and Abstractive-QA.
• Enhanced RAG Performance: Demonstrates improved safety and reduced hallucinations when integrated into RAG systems utilizing Spanish texts.
• Resource-Efficient Training: Achieved its current state in 8.5 hours on a single A100 GPU, leveraging a high-quality dataset and LoRA for memory optimization.
• Competitive Benchmarking: Outperforms several state-of-the-art models, including GPT-4o and Meta-Llama-3.1-8B-Instruct, in specific Spanish RAG-focused evaluation tasks (e.g., AQuAS, RagQuAS, CAM, Shops, Insurance).

Ideal Use Cases

• General Chatbot in Spanish: Functions effectively as a general conversational agent.
• Spanish RAG Systems: Highly recommended for applications requiring accurate information retrieval and generation from Spanish databases.
• Tool Use & Function Calling: Capable of integrating with external tools for enhanced functionality.
• Resource-Constrained Environments: Designed to be usable on hardware with reduced computational capacity, with a GGUF version available for further optimization.