Real-Time Toxicity Filtering for Open-Source Code Reviews
A new AI-powered browser extension filters toxic comments on GitHub with 97% accuracy, fostering healthier open-source collaboration.
A research team has introduced ToxiShield, a novel browser extension designed to combat toxic interactions in open-source development by filtering harmful language in code reviews in real-time. The framework employs a three-module pipeline: first, a fine-tuned BERT-based binary classifier identifies toxic text with a 97% F1-score on a dataset of 38,761 code review texts. For deeper analysis, the system uses Claude 3.5 Sonnet for reasoned multiclass classification (e.g., identifying personal attacks, condescension, or hostility), achieving a 39% Matthews Correlation Coefficient (MCC).
Once toxicity is detected, the core detoxification module, powered by a fine-tuned Llama 3.2 model, rewrites the comment to preserve its technical intent while removing harmful tone. This model demonstrated impressive metrics, including 95.27% style transfer accuracy, 97.03% fluency, and 67.07% content preservation. An initial validation study with 10 software developers indicated that ToxiShield effectively fosters a more inclusive environment. The tool works directly within the browser, meaning developers on platforms like GitHub and GitLab could see toxic comments automatically rephrased into constructive feedback as they browse pull requests and issues.
- Uses a fine-tuned BERT model for toxicity detection with a 97% F1-score on 38,761 code review texts.
- Employs a fine-tuned Llama 3.2 model for detoxification, achieving 95.27% style transfer accuracy and 97.03% fluency.
- Validated by 10 developers, the real-time browser extension aims to reduce friction and improve collaboration in open-source projects.
Why It Matters
This tool directly addresses the human cost of toxic interactions, which can drive contributors away from critical open-source projects.