I just generated 1,000,000 transaction graph visualizations from real Ethereum/Arbitrum/Polygon data — now training a Vision-Language model to detect DeFi attacks

Hey everyone,

I'm building Sigui, a DePIN security oracle for AI agents. Today I hit a milestone I'm proud of:

Dataset: https://huggingface.co/datasets/Ibonon/sigui-depin-1m

What's in it:

• 1,000,000 visual transaction graph images generated from 1.87M real on-chain transactions (Ethereum, Arbitrum, Polygon)
• Each graph is annotated with attack topology labels: DRAIN_STAR, MIXING_CHAIN, NORMAL
• Generated in ~1h15 using 20-core parallel processing on AMD MI300X

What I'm doing with it: I'm currently fine-tuning Qwen2-VL-7B via LoRA on this dataset using AMD ROCm. The goal is a model that sees attack patterns in transaction graphs instead of relying on static rules. This will power Imina-Na V2, the vision brain of my security oracle.

If you want to try V1 right now: https://huggingface.co/Ibonon/imina_na_lora — the first vision model trained on DePIN transaction graphs. Feedback welcome.

The standard behind this: I also co-authored ERC-8259, a proposed Ethereum standard for AI Agent Identity & Threat Registry. https://ethereum-magicians.org/t/erc-8259-ai-agent-identity-threat-registry/28473 https://github.com/ibonon/ERCs

The dataset is fully open (MIT license). Would love feedback on the graph generation approach, annotation quality, or the ERC proposal.