Qwen3-8B-UnBias-Plus-SFT: Specialized Bias Detection and Neutral Rewriting

This model, developed by the Vector Institute, is a fine-tuned version of Qwen3-8B specifically engineered for news media bias detection and neutral rewriting. It operates by analyzing news articles to identify biased language, classify its type and severity, and then generate neutral replacements, ultimately providing a fully rewritten, unbiased version of the article.

Key Capabilities

• Structured JSON Output: Delivers comprehensive analysis in a single JSON response, including binary bias labels, severity scores, and detailed segment-level annotations.
• Segment-Level Bias Detection: Pinpoints exact biased substrings, suggests neutral replacements, and explains the reasoning.
• Bias Type Classification: Identifies various bias types such as loaded language, dehumanizing framing, false generalizations, framing bias, euphemism/dysphemism, politically charged terminology, and sensationalism.
• Automated Neutral Rewriting: Generates a complete, unbiased version of the input article.
• Training: Supervised Fine-Tuning (SFT) with LoRA on the vector-institute/Unbias-plus dataset, which features expert-annotated news articles.

Use Cases and Considerations

This model is particularly well-suited for applications in content moderation, journalistic ethics tools, and automated fact-checking. It is primarily trained on English-language news articles and performs best on articles under 5000 characters. While powerful, human review of outputs is recommended for production use. A smaller, faster 4B parameter version is also available for lighter-weight applications.