GraphWiz/LLaMA2-7B-DPO

GraphWiz/LLaMA2-7B-DPO Overview

GraphWiz/LLaMA2-7B-DPO is a 7 billion parameter instruction-following large language model, built upon the LLaMA 2 architecture, specifically engineered for graph reasoning. It processes natural language descriptions of graphs and structures to explicitly solve various graph problems. The model's training involved a two-stage strategy: Mixed-task Training and DPO Alignment, which significantly enhances its ability to understand and respond to complex graph-related queries.

Key Capabilities & Performance

This model demonstrates strong performance across a range of graph problems, including:

• Cycle Detection: Achieves 89.00% accuracy on cycle detection tasks.
• Connectivity: Scores 82.50% on connectivity problems.
• Bipartite Graph Detection: Reaches 84.75% accuracy.
• Hamiltonian Path/Cycle: Shows a notable 81.50% accuracy.
• Shortest Path & Flow Problems: Achieves 24.00% and 43.50% respectively.

Compared to a Naive SFT LLaMA 2-7B, GraphWiz-DPO shows substantial improvements, particularly in complex reasoning tasks like Hamiltonian paths and flow problems, achieving an average score of 65.00% across various graph tasks. It also outperforms larger models like GPT-4 in several specific graph problem categories.

Good For

• Graph Problem Solving: Ideal for applications requiring an LLM to interpret and solve graph-related challenges from natural language input.
• Research & Development: Useful for researchers exploring advanced reasoning capabilities in LLMs, especially in graph theory and network analysis.
• Educational Tools: Can serve as a powerful backend for tools that teach or demonstrate graph algorithms and concepts.