vector-institute/Qwen3-8B-UnBias-Plus-SFT-Instruct

TEXT GENERATIONConcurrency Cost:1Model Size:8BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:May 25, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Cold

The vector-institute/Qwen3-8B-UnBias-Plus-SFT-Instruct is an 8 billion parameter Qwen3-based language model developed by the Vector Institute, fine-tuned for news media bias detection and neutral rewriting. It identifies biased language segments, classifies bias type and severity, provides neutral replacements, and generates a fully rewritten unbiased version in a single structured JSON response. This model excels at producing clean, structured JSON directly for production inference, making it suitable for applications requiring automated bias analysis and content neutralization.

Loading preview...

Overview

This model, Qwen3-8B-UnBias-Plus-SFT-Instruct, is an 8 billion parameter variant of the Qwen3 architecture, developed by the Vector Institute as part of the UnBias-Plus project. It is specifically fine-tuned for the complex task of news media bias detection and neutralization. Unlike its predecessor, this Instruct variant is trained without chain-of-thought thinking blocks, enabling it to produce clean, structured JSON outputs directly, which is ideal for production environments using vLLM or other OpenAI-compatible backends.

Key Capabilities

  • Comprehensive Bias Analysis: Identifies biased language segments, classifies bias type (e.g., loaded language, dehumanizing framing, sensationalism), and assigns a severity level (0-4).
  • Automated Neutralization: Provides neutral replacement phrases for biased segments and generates a fully rewritten, unbiased version of the input article.
  • Structured JSON Output: Delivers all analysis and rewritten content in a single, parseable JSON format, including severity, bias_found, biased_segments, and unbiased_text.
  • Improved Performance: Achieves a segment localization recall of 0.900 and a segment replacement quality of 3.82/5, with a median latency of 22.2s, outperforming its legacy version.
  • Context Length: Supports a context length of 8192 tokens, allowing for analysis of longer news articles.

Good For

  • Developers building applications for automated news analysis, content moderation, or journalistic tools.
  • Use cases requiring direct, structured output for bias detection and text rewriting without intermediate reasoning steps.
  • Analyzing English-language news articles for various forms of bias, including loaded language, framing bias, and sensationalism.
  • Integration into systems where efficient, production-ready bias detection and neutralization are critical.

Finetunes
2 models