Edaizi/KG-TRACES
Edaizi/KG-TRACES is a 7.6 billion parameter language model developed by Edaizi, designed to enhance explainable reasoning in LLMs through explicit supervision over reasoning paths. It leverages Knowledge Graphs to predict symbolic reasoning paths and generate attribution-aware explanations. This model excels at providing transparent, traceable, and robust reasoning, particularly in specialized fields like medicine, even with limited direct KG access.
Loading preview...
KG-TRACES: Explainable Reasoning with Knowledge Graphs
KG-TRACES is a novel framework that enhances Large Language Models (LLMs) by explicitly supervising their reasoning processes using Knowledge Graphs (KGs). Developed by Edaizi, this 7.6 billion parameter model focuses on making LLM reasoning transparent, accurate, and traceable.
Key Capabilities
- Knowledge Graph-Guided Reasoning: Predicts symbolic reasoning paths from questions to answers using KG guidance.
- Attribution-Aware Explanations: Generates reasoning explanations that clearly indicate whether each step originates from the KG or the LLM's internal knowledge.
- Robust Performance: Maintains strong performance even when direct KG access is limited or unavailable during inference.
- Versatile Application: Demonstrates strong generalization capabilities, including in specialized domains like medicine.
How it Works
KG-TRACES teaches LLMs to reason by guiding them to chart knowledge graph reasoning paths and show their work through attribution-aware explanations. This approach ensures crystal-clear explanations and trustworthy, attributable reasoning.
Datasets
The model utilizes augmented SFT datasets for WebQSP and CWQ, which include reasoning paths and augmented reasoning processes with source attributions. These datasets are available on Hugging Face: KG-TRACES-WebQSP and KG-TRACES-CWQ.
Good for
- Applications requiring explainable and verifiable AI reasoning.
- Use cases where attribution of information sources is critical.
- Domains demanding robust reasoning even with varying access to external knowledge bases.