vector-institute/Qwen3-8B-UnBias-Plus-SFT
Qwen3-8B-UnBias-Plus-SFT is an 8 billion parameter language model developed by the Vector Institute, fine-tuned from Qwen3-8B using Supervised Fine-Tuning (SFT) with LoRA. This model specializes in news media bias detection, classification, and neutral rewriting, providing structured JSON outputs for identified biased language segments and fully rewritten unbiased articles. It features an 8192-token context length and is optimized for analyzing English-language news content.
Loading preview...
Overview
vector-institute/Qwen3-8B-UnBias-Plus-SFT is a specialized 8 billion parameter language model, fine-tuned by the Vector Institute from the Qwen3-8B base model. It was developed as part of the UnBias-Plus project, focusing on the critical task of news media bias detection and mitigation. The model is trained to identify biased language within news articles, classify the type and severity of bias, suggest neutral replacements, and ultimately generate a completely rewritten, unbiased version of the input article.
Key Capabilities
- Bias Detection: Identifies specific biased language segments within news articles.
- Bias Classification: Categorizes detected biases into types such as loaded language, dehumanizing framing, false generalizations, framing bias, euphemism/dysphemism, politically charged terminology, and sensationalism.
- Severity Assessment: Assigns a severity score (0, 2, 3, 4) to the overall article and individual biased segments, indicating neutrality, recurring biased framing, strong persuasive tone, or inflammatory rhetoric.
- Neutral Rewriting: Provides neutral alternative phrases for biased segments and generates a fully rewritten, unbiased version of the entire article.
- Structured Output: Delivers all analysis and rewritten content in a single, structured JSON response, including a binary bias label, overall severity, and detailed information for each biased segment.
- Context Length: Supports an 8192-token context, allowing for analysis of substantial news articles.
Training and Usage
The model was fine-tuned using Supervised Fine-Tuning (SFT) with LoRA on the UnBias-Plus (train_1) dataset. It operates with bf16 precision and can be run with approximately 16GB VRAM (full precision) or 5GB VRAM (4-bit quantization). While powerful, it is primarily trained on English-language news articles and performs best on articles under 5000 characters. Human review of outputs is recommended for production use.