Overview
This model, Qwen3-8B-UnBias-Plus-SFT-Instruct, is an 8 billion parameter instruction-tuned variant of Qwen3-8B, developed by the Vector Institute as part of the UnBias-Plus project. It is specialized in analyzing news articles to detect biased language, classify bias types and severity, suggest neutral replacements, and generate a fully rewritten unbiased version of the article. The model is fine-tuned using Supervised Fine-Tuning (SFT) with LoRA on the vector-institute/Unbias-plus dataset, which contains expert-annotated news articles.
Key Capabilities
- Comprehensive Bias Detection: Identifies biased language segments within news articles.
- Bias Classification: Classifies detected bias by type (e.g., loaded language, dehumanizing framing, sensationalism) and severity (Neutral, Recurring biased framing, Strong persuasive tone, Inflammatory rhetoric).
- Neutral Rewriting: Provides neutral alternative phrases for biased segments and generates a complete unbiased version of the input article.
- Structured JSON Output: Delivers all analysis and rewritten content in a single, structured JSON response, making it easy for programmatic integration.
- Production Optimized: Unlike its SFT counterpart, this Instruct variant disables chain-of-thought thinking (
enable_thinking=False), leading to lower latency and direct JSON output, making it suitable for production APIs and vLLM deployment.
Good For
- News Media Analysis: Automating the detection and mitigation of bias in news content.
- Content Moderation: Identifying and addressing biased language in various textual data.
- API Deployment: Ideal for integration into applications requiring fast, reliable, and structured bias analysis via vLLM or other OpenAI-compatible serving backends.
- Research: Studying and understanding linguistic bias patterns, particularly in English-language news.