GraphWiz/LLaMA2-13B-DPO

GraphWiz/LLaMA2-13B-DPO: Graph Reasoning LLM

GraphWiz/LLaMA2-13B-DPO is a 13 billion parameter instruction-following language model built upon the LLaMA 2 architecture, developed by GraphWiz. Its core capability lies in understanding natural language descriptions of graphs and structures, and subsequently solving complex graph problems. This model is distinguished by its specialized training, which involves two key stages: Mixed-task Training and DPO Alignment, specifically tailored to enhance its graph reasoning abilities.

Key Capabilities

• Graph Problem Solving: Proficiently solves various graph-related tasks, including cycle detection, connectivity analysis, bipartite graph determination, topology identification, shortest path finding, triangle counting, flow problems, Hamiltonian path detection, and subgraph identification.
• Natural Language Interaction: Maps textual graph descriptions and provides solutions explicitly in natural language, making it accessible for users to query graph properties and solutions.
• Enhanced Performance: Demonstrates significant improvements over naive SFT (Supervised Fine-Tuning) versions of LLaMA 2-13B and Mistral-7B on a suite of graph reasoning benchmarks, achieving an average score of 63.89% on the evaluated tasks.

Good For

• Graph Analysis: Ideal for applications requiring an LLM to interpret and reason about graph data presented in text.
• Educational Tools: Can be used to explain graph theory concepts and solutions in an intuitive, natural language format.
• Research in Graph Neural Networks and LLMs: Serves as a strong baseline or component for further research into combining large language models with graph reasoning capabilities.

For more details, refer to the project page and the accompanying paper.