Can Small Models Beat Frontier LLMs? Introducing SuiAudit Pure1: Specialized Auditing via Extreme Low-Resource Fine-Tuning

The rapid proliferation of Large Language Models (LLMs) has revolutionized automated code auditing, yet frontier models like GPT-5.2 and Claude Opus 4.5 continue to struggle with niche, high-performance languages such as Sui Move. These generalist models frequently exhibit "Solidity Bias," hallucinating Ethereum-specific vulnerabilities (e.g., reentrancy) in Sui's objectcentric architecture. Furthermore, their reliance on cloud-based inference compromises developer privacy and intellectual property. To address these limitations, we introduce SuiAudit Pure1, a specialized 7B-parameter model explicitly engineered for two primary objectives: (1) enabling secure, real-time offline inference suitable for local development environments (e.g., VS Code plugins) to protect intellectual property, and (2) eliminating the "Solidity Bias" by mastering Sui Move's specific safety guarantees (e.g., absence of reentrancy or overflow issues). Our distinct contribution lies in our dataset design, which balances positive vulnerability instances (Classes A-F) against adversarial "Hard Negative" samples (Class NEG)-safe code patterns designed to trap generalist LLMs. By employing a progressive training pipeline, we first establish a robust baseline via Supervised Fine-Tuning (SuiAudit SFT) to master domain syntax, followed by Direct Preference Optimization (DPO) with Chain-of-Thought to refine reasoning. This yields our final SuiAudit Pure1 model, which achieves an accuracy of 31.4% on our rigorous benchmark, significantly surpassing Claude Opus 4.5 (26.6%) and GPT-5.2 (17.2%). To facilitate reproducibility while protecting intellectual property, we released our introductory SuiAudit Raw1 infrastructure on GitHub. We provide open access to the model weights for both stages of our pipeline: the intermediate SuiAudit SFT baseline and the fully aligned SuiAudit Pure1. The proprietary dataset used for training remains a closed asset.