mourningdove/zk-auditor

TEXT GENERATIONConcurrency Cost:1Model Size:1.5BQuant:BF16Ctx Length:32kPublished:Mar 26, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Cold

mourningdove/zk-auditor is a LoRA fine-tune of Qwen2.5-Coder-1.5B-Instruct, specifically designed to identify insufficient constraints in Zero-Knowledge (ZK) proof circuits. This 1.5 billion parameter model is specialized for security auditing of Circom circuits, serving as a proof-of-concept for automated vulnerability detection. Its primary use case is first-pass triage during security audits and as an educational tool for understanding constraint insufficiency patterns in ZK proofs.

Loading preview...

ZK Constraint Auditor v0.0

mourningdove/zk-auditor is a specialized LoRA fine-tune of the Qwen2.5-Coder-1.5B-Instruct model, focusing on the critical task of identifying insufficient constraints within Zero-Knowledge (ZK) proof circuits, particularly those written in Circom. This initial version, while a proof of concept, demonstrates the potential for AI-assisted security auditing in the ZK space.

Key Capabilities

  • Specialized ZK Circuit Auditing: Fine-tuned to detect common patterns of constraint insufficiency in Circom circuits.
  • Security Triage: Intended for initial vulnerability assessment during security audits, providing a first-pass analysis.
  • Educational Tool: Can assist in learning and understanding typical constraint-related vulnerabilities in ZK proofs.
  • MLX Framework Support: Optimized for use with the MLX framework, with usage examples provided for macOS.

Good for

  • Security Auditors: Performing preliminary checks on Circom circuits to quickly identify potential weaknesses.
  • ZK Developers: Learning about common security pitfalls and ensuring robust constraint design in their circuits.
  • Researchers: Establishing a baseline for further research into AI-driven ZK security tooling.

This model was trained on a limited dataset (~15 examples) from the mourningdove007/zk-constraint-data repository, making it most effective on very simple circuits. Future versions (v0.1, v0.2) aim to expand the training data with real findings from public audits for broader pattern coverage and improved performance.

Finetunes
1 models