Offlin33er/qwen25-coder-7b-solidity-audit-v3-merged

TEXT GENERATIONPricing:Input $0.4 / Cached $0.08 / Output $0.8Concurrent Unit Cost:1Model Size:7.6BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 26, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

Offlin33er/qwen25-coder-7b-solidity-audit-v3-merged is a 7.6 billion parameter Qwen2.5-Coder-7B-Instruct model fine-tuned by Offlin33er, specifically designed for Solidity smart contract security auditing. This model excels at generating structured audit reports, classifying findings into one of eight canonical vulnerability categories. It provides functionality analysis, vulnerability descriptions, and remediation recommendations for Solidity code snippets, making it suitable for automated security analysis.

Loading preview...

Model Overview

This model, Offlin33er/qwen25-coder-7b-solidity-audit-v3-merged, is a full-weight merge of a QLoRA adapter into the Qwen/Qwen2.5-Coder-7B-Instruct base model. With 7.6 billion parameters and a context length of 32768 tokens, it is specifically engineered for Solidity smart contract security auditing. It is part of the larger Solidity Audit Scanner project and is optimized for deployment on Hugging Face Inference Endpoints.

Key Capabilities

  • Structured Audit Reports: Given a Solidity code snippet, the model produces a detailed audit report.
  • Vulnerability Classification: It classifies primary findings into one of eight canonical categories: Reentrancy, Arithmetic Error, Access Control Flaw, Input Validation Error, Frontrunning / MEV, Denial of Service (DoS), Business Logic Violation, or Clean (No Vulnerability).
  • Detailed Analysis: Reports include functionality analysis, vulnerability descriptions, and remediation recommendations.
  • High Detection Precision: Achieves a detection precision of 0.986 and recall of 0.986 on a stratified held-out evaluation set.

Training and Performance

The model was trained for one epoch (139 steps) on the Offlin33er/solidity-audit-instruct-v3 dataset, which contains 2,224 rows with an 8-class canonical taxonomy. While detection metrics are strong, the exact finding-type match is 0.449, indicating some semantic ambiguity in reference labels rather than model bugs. It is recommended for deployment on an Nvidia L4 GPU (24 GB) for optimal performance.

Limitations

  • Trained on approximately 2.5K Solidity-audited snippets.
  • Not a substitute for professional human audits; findings should be considered candidates for human review.
  • Around half of detected findings may receive an adjacent canonical label due to reference-label ambiguity.

Finetunes
1 models