This paper presents research to identify vulnerabilities, biases, and malicious components in the supply chain of large-scale language models (LLMs), improve model fairness, and ensure compliance with regulatory frameworks. Given that existing LLMs inevitably address these issues due to their reliance on base models, pre-trained models, and external datasets, we study the LLM supply chain, focusing on the relationships between models and datasets. To this end, we design a methodology to systematically collect LLM supply chain information and construct a novel directed heterogeneous graph (402,654 nodes and 462,524 edges) representing the relationships between models and datasets. This graph is then used to perform various analyses and yield several interesting results.
