vector-institute/Qwen3-8B-UnBias-Plus-SFT
Qwen3-8B-UnBias-Plus-SFT is a 8 billion parameter language model developed by the Vector Institute, fine-tuned from Qwen3-8B. This model specializes in news media bias detection and neutral rewriting, providing structured JSON outputs that identify biased language, classify bias types and severity, and offer unbiased replacements. It is designed to return a fully rewritten, unbiased version of an article, making it ideal for applications requiring automated content neutrality analysis.
Loading preview...
Overview
Qwen3-8B-UnBias-Plus-SFT is a specialized 8 billion parameter language model developed by the Vector Institute, fine-tuned from the Qwen3-8B base model. Its core function is to analyze news articles for biased language, classify the type and severity of bias, and generate neutral replacements, ultimately providing a fully rewritten, unbiased version of the input article. The model outputs a structured JSON response containing detailed information on detected biases and the neutral text.
Key Capabilities
- Bias Detection: Identifies biased language segments within news articles.
- Bias Classification: Classifies bias into types such as loaded language, dehumanizing framing, false generalizations, and more, along with severity levels (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 article.
- Structured Output: Delivers results in a consistent JSON format, including a binary bias label, overall severity, and details for each biased segment.
- Context Length: Supports a context length of 8192 tokens, allowing for analysis of substantial articles.
Training and Usage
The model was fine-tuned using Supervised Fine-Tuning (SFT) with LoRA on the UnBias-Plus (train_1) dataset. It supports both full precision (bf16) for server environments and 4-bit quantization for lightweight deployment on devices with limited VRAM (e.g., laptops).
Limitations
- Primarily trained on English-language news articles.
- Political bias detection reflects patterns present in the training data.
- Best performance is observed on articles under 5000 characters.
- Human review of outputs is recommended for production use.