vector-institute/Qwen3-8B-UnBias-Plus-SFT-Instruct-V2

TEXT GENERATIONConcurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 13, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

vector-institute/Qwen3-8B-UnBias-Plus-SFT-Instruct-V2 is an 8 billion parameter Qwen3-based 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 V2 Instruct variant is specifically designed for production inference, directly producing clean structured JSON without chain-of-thought thinking blocks. It acts as a conservative span-level bias annotator, flagging only clear, material bias and applying strict label precedence.

Loading preview...

Overview

vector-institute/Qwen3-8B-UnBias-Plus-SFT-Instruct-V2 is an 8 billion parameter model, fine-tuned by the Vector Institute for news media bias detection and neutral rewriting. Built upon the Qwen3-8B base, this model is designed to analyze news articles, identify biased language, classify its type and severity, and provide neutral replacements, ultimately generating a fully rewritten unbiased version. It operates as a conservative span-level bias annotator, focusing on clear, material bias and adhering to strict bias-type label precedence.

Key Capabilities

  • Structured JSON Output: Directly produces clean, structured JSON responses, making it suitable for production environments and vLLM inference.
  • Comprehensive Bias Analysis: Identifies biased segments, assigns one of eight canonical bias types (e.g., dehumanizing_language, sensationalism, opinion_as_fact), and provides a segment-level severity (Low, Medium, High).
  • Neutral Rewriting: Generates an unbiased_text version of the article, changing only the flagged biased text.
  • Article-level Severity: Assigns an overall article severity score from 0-10.
  • Enhanced Training: Trained on train_4, a dataset of 5,000 expert-annotated news articles, an improvement over its predecessor's train_3 dataset.

What Makes This Model Different (V2 Instruct)

This V2 Instruct variant is specifically trained without chain-of-thought thinking blocks (enable_thinking=False), ensuring direct and clean structured JSON output. It features an updated bias type taxonomy with 8 canonical slugs and strict label precedence, a refined article severity scale (0-10 integer), and a system prompt tailored for conservative span-level annotation, including rules for attribution handling and toxicity. Its focus on direct, structured output and conservative annotation distinguishes it for applications requiring precise, actionable bias analysis.

Finetunes
1 models