"Optimal Cognitive Core"- specialized 1.7B model for grounded question answering
14 hours ago
- #Question Answering
- #Retrieval-Augmented Generation
- #AI Model
- Instructions for using occ-ai/OCC-RAG-1.7B with various libraries, notebooks, and local apps.
- Model is a 1.7B-parameter specialized for faithful, context-grounded QA with structured reasoning.
- Highlights: faithfullness, calibrated abstention, structured reasoning, and compact size.
- Evaluation shows strong performance in multi-hop reasoning, faithfulness, and refusal accuracy.
- Input/output uses structured prompts with special tokens and sections for query, source analysis, reasoning, status, and answer.
- Quickstart example with Transformers shows how to use the model with documents and parse responses.
- Deployment compatible with vLLM, SGLang, and other serving stacks, suitable for constrained infrastructure.
- Limitations include context-grounded only, capped reasoning depth at three hops, and not a general-purpose model.