Overview

ProductsLlama is an 8 billion parameter language model developed by xinyifang, specifically designed for node classification on text-attributed graphs (TAGs). Fine-tuned from Llama3.1_8b_Ins_Arxiv_4096, this model addresses challenges like input window size limitations and computational overhead in large-scale graph processing.

Key Capabilities

• Optimized for Text-Attributed Graphs: Excels at classifying nodes within graphs where nodes have associated textual information, such as product descriptions.
• Multi-Profiling Framework: Utilizes a novel data augmentation method involving multiple profiling/summarizing models to increase training data diversity and quantity.
• Efficient Graph Processing: Designed to process complex graphs efficiently, even with limited computational resources, by constructing concise yet informative fine-tuning prompts.
• Strong Performance: Outperforms 11 state-of-the-art baselines, achieving 74.31% accuracy on ogbn-arxiv and 85.15% accuracy on ogbn-products datasets.
• Fast Training: The model was trained 2x faster using Unsloth and Hugging Face's TRL library.

Good For

• Product Categorization: Accurately classifying product descriptions into predefined categories.
• Node Classification: Performing classification tasks on text-attributed graphs in domains like social networks and recommender systems.
• Research in Graph Neural Networks: Serving as a strong baseline or component for further research in LLM applications for graph data.

This model is based on the paper "LLM Profiling and Fine-Tuning with Limited Neighbor Information for Node Classification on Text-Attributed Graphs" (IEEE Xplore).