---
license: mit
language:
- en
tags:
- blockchain
- defi-security
- depin
- graph
- vision
- transaction-analysis
- ethereum
- arbitrum
- polygon
- attack-detection
pretty_name: Sigui DePIN 1M — Multichain Transaction Graph Dataset
size_categories:
- 1M<n<10M
task_categories:
- image-classification
- image-text-to-text
---
# Sigui DePIN 1M — Multichain Transaction Graph Dataset
> The largest open dataset of annotated blockchain transaction graph visualizations for AI security research.
---
## 📋 Dataset Description
**sigui-depin-1m** contains 1,000,000 visual transaction graph images generated from 1.87 million real on-chain transactions across Ethereum, Arbitrum, and Polygon. Each image is annotated with an attack topology label designed to train Vision-Language Models to detect DeFi threats.
This dataset is the training foundation for **Imina-Na V2**, the vision brain of the Sigui Protocol — a DePIN security oracle for the agentic economy.
---
## 📊 Dataset Statistics
| Property | Value |
|---|---|
| **Total images** | 1,000,000 |
| **Source transactions** | 1,870,000+ real on-chain txs |
| **Chains covered** | Ethereum, Arbitrum, Polygon |
| **Label classes** | 3 (DRAIN_STAR, MIXING_CHAIN, NORMAL) |
| **Image format** | PNG (dark background) |
| **Total size** | ~10.8 GB (images.tar) |
| **Annotation file** | qwen2_vl_real_data.jsonl (260 MB) |
| **License** | MIT |
---
## 🏷️ Label Classes
### DRAIN_STAR
Multiple wallets converging simultaneously into a single target address. This topology is the signature of coordinated rug pulls, flash loan attacks, and fund aggregation schemes.
### MIXING_CHAIN
Funds routed through a sequential chain of intermediate wallets to obscure the origin. This is the on-chain fingerprint of mixer evasion, layering attacks, and money laundering.
### NORMAL
Organic transaction topologies with no anomalous structure. Standard user-to-user and contract interactions.
---
## 📁 Dataset Structure
sigui-depin-1m/ ├── images.tar # 1,000,000 PNG graph images └── qwen2_vl_real_data.jsonl # Annotations in ShareGPT format

Each line in the JSONL file follows this format:
```json
{
  "messages": [
    {"role": "user", "content": "<image>Analyze this DePIN transaction graph."},
    {"role": "assistant", "content": "DRAIN_STAR detected."}
  ],
  "images": ["/path/to/graph_000001_drain_star.png"]
}
⚙️ Generation Process
Generated using a custom multiprocessing pipeline on AMD MI300X (ROCm 7.0):

Source: 1.87M real transactions from Ethereum, Arbitrum, and Polygon via public APIs
Graph construction: NetworkX DiGraph built from from/to transaction pairs
Attack injection: Deterministic injection of DRAIN_STAR and MIXING_CHAIN patterns
Rendering: Matplotlib (Agg backend), dark background, DPI=50
Speed: ~225 images/second using 20 parallel CPU cores
Total generation time: ~1 hour 15 minutes
🔧 How to Use
python
from datasets import load_dataset
dataset = load_dataset(
    "json",
    data_files="qwen2_vl_real_data.jsonl",
    split="train"
)
print(dataset[0])
🔗 Related Resources
Sigui Protocol (GitHub): https://github.com/ibonon/ERCs
ERC-8259 Standard: https://ethereum-magicians.org/t/erc-8259-ai-agent-identity-threat-registry/28473
Imina-Na V1 Model: https://huggingface.co/Ibonon/imina_na_lora
Imina-Na V2 Model: https://huggingface.co/Ibonon/imina_na_v2_lora
✍️ Citation
bibtex
@dataset{sigui_depin_1m_2026,
  author    = {Ibonon},
  title     = {Sigui DePIN 1M — Multichain Transaction Graph Dataset},
  year      = {2026},
  publisher = {Hugging Face},
  url       = {https://huggingface.co/datasets/Ibonon/sigui-depin-1m}
}
⚖️ License
MIT — Open for research and commercial use.