HuggingGraph: Understanding the Supply Chain of LLM Ecosystem

---

📖 Overview

HuggingGraph is a heavy-tailed, heterogeneous graph that captures Supply Chain Relationships across Large Language Models (LLMs) and datasets. It is constructed from Hugging Face metadata, repository cross-links, and textual analysis, modeling how datasets, base models, fine-tunes, adapters, quantized models, and merges interconnect in the AI supply chain.

This repository contains the artifacts released with our CIKM 2025 paper:

HuggingGraph: Understanding the Supply Chain of the LLM Ecosystem

---

📂 Repository Contents

• HuggingGraph.dot – Full graph in Graphviz .dot format.
• subgraph.pdf – A sample subgraph figure (as shown in the paper), highlighting the supply chain relationship for a representative model.
• README.md – Documentation and usage instructions.

---

⚡ Getting Started

To work with HuggingGraph, you will need:

1. Install Graphviz

Graphviz is required to render the .dot graph into PDF, PNG, or SVG formats.

# Ubuntu / Debian
sudo apt-get install graphviz

# macOS (Homebrew)
brew install graphviz

# Windows (Chocolatey)
choco install graphviz

Verify installation:

dot -V

---

2. Set Up Python Environment

We recommend Python 3.10+. Create a virtual environment:

python -m venv venv
source venv/bin/activate   # Linux / macOS
venv\Scripts\activate      # Windows

Install dependencies:

pip install networkx pydot

---

🚀 Usage

1. Access the Graph

The repository now directly provides the graph as a Graphviz DOT file:

• HuggingGraph.dot → the full constructed graph of the LLM ecosystem
• subgraph.pdf → a smaller excerpt for quick visualization and presentations

No extraction step is needed — you can work with the .dot file directly.

---

2. Visualize with Graphviz

You can render the .dot file into different formats (e.g., SVG, PNG, PDF) using Graphviz.

Render the graph:

# Render to SVG
dot -Tsvg HuggingGraph.dot -o HuggingGraph.svg

# Render to PDF
dot -Tpdf HuggingGraph.dot -o HuggingGraph.pdf

# Render to PNG
dot -Tpng HuggingGraph.dot -o HuggingGraph.png

⚠️ Directly rendering the full graph may be impractical due to its size. Start with subgraphs or use the provided subgraph.pdf for illustration.

---

3. Programmatic Exploration (Python / NetworkX)

You can load and analyze the graph with NetworkX:

import networkx as nx
from networkx.drawing.nx_pydot import read_dot

# Load the graph
G = read_dot("HuggingGraph.dot")

# Print basic statistics
print(f"Nodes: {G.number_of_nodes()}")
print(f"Edges: {G.number_of_edges()}")

# Example: list first 10 nodes
print(list(G.nodes())[:10])

---

4. Working with Subgraphs

Because the full graph is very large, it is often more practical to work with smaller subgraphs.

Example: extract the first 100 nodes as a subgraph:

subset_nodes = list(G.nodes())[:100]
H = G.subgraph(subset_nodes)

# Export to DOT format
nx.nx_pydot.write_dot(H, "subgraph.dot")

Then render with Graphviz as before:

dot -Tpdf subgraph.dot -o subgraph.pdf

Or simply open the provided subgraph.pdf for a ready-to-use illustration.

---

📊 Graph Schema

• Node types:

  • dataset, base, finetune, adapter, quantization, merge

• Edge semantics:

  • Model → Model: derivations (fine-tune, adapter, quantization, merge)
  • Dataset → Model / Model → Dataset: training or usage links
  • Dataset → Dataset: subset or variant relationships

---

📑 Paper Context

HuggingGraph enables systematic analysis of the LLM ecosystem:

• Scale: over 400K nodes and 460K edges.

• Findings:

  • Ecosystem is hub-dominated — a few datasets (e.g., The Pile) and base models (e.g., LLaMA) influence thousands of downstream artifacts.
  • Daily churn — fine-tunes and variants emerge at high velocity, reshaping the graph continuously.
  • Risks — poor provenance, missing metadata, and hidden dataset reuse introduce vulnerabilities in security, bias, and licensing.

HuggingGraph supports forward and backward tracing of dependencies, helping researchers, auditors, and policymakers validate provenance and detect inherited risks.

---

📎 Citation

If you use this dataset or figures, please cite the paper:

# Citation details will be updated soon

---

📬 Contact

For questions, updates, and related work, visit:
👉 Yuede Ji
👉 Mohammad Shahedur Rahman