"Optimal Cognitive Core"- specialized 1.7B model for grounded question answering (huggingface.co)

The announcement of the OCC-RAG-1.7B model introduces a specialized 1.7 billion-parameter language model designed for context-grounded question answering. This model, part of the Optimal Cognitive Core (OCC) series, excels in producing structured reasoning traces with explicit source citations, enabling it to determine the extent to which context supports an answer and to abstain when necessary. Notably, despite its smaller size, OCC-RAG-1.7B surpasses larger general-purpose models in multi-hop reasoning and faithfulness benchmarks, showcasing the ability to provide accurate and reliable citations for its answers. This development is a game-changer for the AI/ML community, particularly in fields where accountability and transparency in AI responses are critical. OCC-RAG-1.7B achieves superior performance in faithfulness metrics, exhibiting the lowest memorization ratio compared to other models, which helps avoid misinformation. The model's innovative training process, utilizing a large synthetic corpus of multi-context QA, along with a structured prompt format, allows for efficient deployment in resource-constrained environments like desktop systems. As AI systems increasingly integrate into decision-making processes, OCC-RAG-1.7B’s approach to truthfulness and reasoning could significantly enhance the reliability of automated services in various applications.