vector-institute/Qwen3.5-4B-UnBias-Plus-SFT-Instruct

VISIONConcurrent Unit Cost:1Model Size:4.5BQuant:BF16Context Size:32kTool Calling:SupportedPublished:May 29, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

The vector-institute/Qwen3.5-4B-UnBias-Plus-SFT-Instruct is a 4.5 billion parameter instruction-tuned causal language model developed by the Vector Institute. Fine-tuned from Qwen3.5-4B, this model specializes in detecting news media bias, classifying bias types and severity, and generating neutral rewrites in a structured JSON format. It is optimized for speed and efficient bias detection without chain-of-thought processing, making it suitable for real-time news analysis and content moderation. The model achieves near-perfect rewrite length preservation and zero duplicate segment rate, trained on 3,000 expert-annotated news articles.

Loading preview...

Overview

This model, Qwen3.5-4B-UnBias-Plus-SFT-Instruct, is a specialized 4.5 billion parameter language model developed by the Vector Institute. It is fine-tuned from Qwen3.5-4B using Supervised Fine-Tuning (SFT) with LoRA, specifically for news media bias detection and neutral rewriting.

Key Capabilities

  • Bias Detection and Rewriting: Given a news article, it identifies biased language segments, classifies their type and severity, provides neutral replacements, and returns a fully rewritten unbiased version.
  • Structured JSON Output: All outputs are provided in a structured JSON format, including overall severity, bias_found flag, biased_segments with original, replacement, severity, bias_type, and reasoning, and the unbiased_text.
  • Specific Bias Types: Detects various bias types such as loaded language, dehumanizing framing, false generalizations, framing bias, euphemism/dysphemism, politically charged terminology, and sensationalism.
  • Optimized for Speed: This "Instruct variant" is trained without chain-of-thought thinking (enable_thinking=False), making it the fastest model in the UnBias-Plus family.
  • Performance Metrics: Achieves a 99.1% parse rate, 57.5% bias reduction, 0.0% duplicate rate, and a median length ratio of 1.000, indicating excellent rewrite length preservation.

Use Cases

  • News Media Analysis: Ideal for automated analysis of news articles to identify and mitigate bias.
  • Content Moderation: Can assist in ensuring neutrality in published content.
  • Research: Useful for researchers studying media bias and natural language processing applications in this domain.

Limitations

  • Primarily trained on English-language news articles.
  • Context length is limited to 4096 tokens.
  • Outputs should be human-reviewed before production use.

Finetunes
1 models