TroyDoesAI/Phi-3-Context-Obedient-RAG

Overview

This model, TroyDoesAI/Phi-3-Context-Obedient-RAG, is a specialized fine-tune of Microsoft's Phi-3-mini-128k-instruct base model. Its core purpose is to significantly enhance context adherence and mitigate hallucinations, particularly within Retrieval Augmented Generation (RAG) workflows. The model achieves this through a unique, structured prompt format that explicitly delineates input blocks, associated context (metadata), and instructions.

Key Capabilities

• Enhanced Context Adherence: Designed to strictly follow the provided contextual information, reducing the generation of ungrounded responses.
• Reduced Hallucinations: By enforcing context obedience, the model aims to minimize instances where it generates information not present in the source material.
• Precise Source Referencing: The training dataset includes examples where the model is prompted to cite specific source details (e.g., date, URL) from the provided context in its responses, addressing a common weakness in RAG systems.
• Structured Prompting: Utilizes explicit delimiters (BEGININPUT, BEGINCONTEXT, ENDCONTEXT, ENDINPUT, BEGININSTRUCTION, ENDINSTRUCTION) to clearly define different parts of the prompt, helping the model accurately locate and utilize information.

Good For

• RAG Applications: Ideal for scenarios where grounding responses in retrieved documents is critical.
• Fact-Checking & Verification: Useful for systems requiring verifiable answers linked directly to source material.
• Reducing AI Hallucinations: Aims to improve the trustworthiness of AI-generated content by ensuring it stays within the bounds of provided facts.
• Complex Information Retrieval: When dealing with multiple document chunks, this model can help attribute specific parts of the answer to the correct source.