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

TEXT GENERATIONConcurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:May 25, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Qwen3-8B-UnBias-Plus-SFT-Instruct is an 8 billion parameter instruction-tuned causal language model developed by the Vector Institute. Fine-tuned from Qwen/Qwen3-8B, it specializes in detecting and neutralizing news media bias, providing structured JSON outputs for biased segments, their types, severity, and a fully rewritten unbiased article. This model is optimized for production inference, directly generating clean JSON without intermediate thinking steps, and has a context length of 8192 tokens.

Loading preview...

Model Overview

Qwen3-8B-UnBias-Plus-SFT-Instruct is an 8 billion parameter language model developed by the Vector Institute as part of the UnBias-Plus project. It is a fine-tuned variant of Qwen/Qwen3-8B, specifically designed for news media bias detection and neutral rewriting. Unlike its predecessor, this "Instruct variant" is trained to directly output structured JSON responses without chain-of-thought reasoning, making it suitable for efficient production deployment.

Key Capabilities & Features

  • Bias Detection and Neutralization: Identifies biased language segments in news articles, classifies bias type and severity, and suggests neutral replacements.
  • Structured JSON Output: Provides a comprehensive JSON response including severity, bias_found, biased_segments (with original, replacement, severity, bias_type, reasoning), and unbiased_text.
  • Specific Bias Types: Detects loaded language, dehumanizing framing, false generalizations, framing bias, euphemism/dysphemism, politically charged terminology, and sensationalism.
  • Optimized for Inference: Designed for direct JSON output (enable_thinking=False) for use with vLLM or other OpenAI-compatible backends.
  • 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.
  • Context Length: Supports an 8192-token context, allowing for analysis of longer articles.

Use Cases & Limitations

This model is ideal for applications requiring automated, structured analysis and rewriting of news content to mitigate bias. It excels at identifying and replacing biased phrases, providing a fully rewritten neutral article. However, it is primarily trained on English-language news, and its handling of already unbiased articles is less robust than the Legacy model. Outputs should always be human-reviewed before production use, especially for articles exceeding 5,000 characters or those with potential rewrite expansion issues.

Finetunes
3 models