GraphWiz/LLaMA2-7B
GraphWiz/LLaMA2-7B is a 7 billion parameter instruction-following large language model developed by GraphWiz, built upon the LLaMA 2 architecture with a 4096 token context length. This model is specifically fine-tuned to interpret textual descriptions of graphs and structures, and then solve various graph-related problems expressed in natural language. It excels at graph reasoning tasks, demonstrating improved performance on problems like cycle detection, connectivity, and shortest path calculations compared to base LLaMA 2 models.
Loading preview...
GraphWiz/LLaMA2-7B: Graph Reasoning LLM
GraphWiz/LLaMA2-7B is an instruction-following large language model (LLM) specifically designed to understand and solve graph problems described in natural language. Built on the LLaMA 2-7B architecture, it undergoes a two-stage training process: Mixed-task Training and DPO Alignment, which significantly enhances its graph reasoning capabilities.
Key Capabilities
- Natural Language Graph Interpretation: Processes textual descriptions of graphs and structures.
- Graph Problem Solving: Solves various graph-related tasks, including cycle detection, connectivity, bipartite graph identification, topology analysis, shortest path, triangle counting, flow problems, Hamiltonian paths, and subgraph detection.
- Enhanced Performance: Benchmarks show that GraphWiz-DPO (LLaMA 2-7B) achieves an average accuracy of 65.00% across diverse graph tasks, outperforming naive SFT LLaMA 2-7B (44.81%) and even some larger models like GPT-4 in specific categories.
- Instruction Following: Designed to follow instructions for complex graph queries.
Good For
- Graph-based Reasoning: Ideal for applications requiring an LLM to reason about and solve problems on graphs described in text.
- Research and Development: Useful for exploring the intersection of large language models and graph theory, particularly for tasks involving structured data interpretation.
- Automated Graph Analysis: Can be leveraged for automating the analysis of graph properties from natural language inputs.