Technical Consideration

Base Model Selection

The following coding-focused models will be tested to find the best performed one:

• CodeLlama-70B (Meta)

• DeepSeek-Coder

• Mistral Codestral

• Qwen 2.5 Coder

• Claude 3.7

Vertical Approach

Targeted Fine-Tuning

• Use Solidity/EVM-specific datasets (high-quality GitHub contracts + security audit reports)

• Incorporate vulnerability pattern datasets

• Include other programming language datasets (Rust, Move, etc.)

Knowledge Enhancement

• Integrate OpenZeppelin standard libraries and EIP documentation as RAG (Retrieval-Augmented Generation) knowledge sources

Data Strategy

High-Quality Data Sources

Positive Samples

• Audited, well-known protocol code (e.g., Uniswap V4, Compound)

• Standard implementations from OpenZeppelin Contracts

• Reference implementations from EIP proposals (ERC-20/721/4337)

Negative Samples

• Smart contract vulnerability case libraries (e.g., DASP Top 10)

• Reconstructed code from DeFi attacks (The DAO, Parity Multisig, etc.)

• Edge cases generated by fuzz testing

Priority Smart Contract Types for Training

• Token Contracts

• Payment Contracts

• Governance Contracts

• Identity Contracts

• Staking Contracts

• Lending Contracts

• Exchange Contracts

• Bridge Contracts

• Multisig Contracts

• Prediction Market Contracts