Hugging Face Overview
AuthorMist Originality is a specialized 3.1 billion parameter language model, built upon Qwen2.5-3B Instruct, designed to make AI-generated text indistinguishable from human-written content. It achieves this by preserving the original meaning while altering the writing style to bypass AI text detection. The model was fine-tuned using Group Relative Policy Optimization (GRPO) with detector feedback as a reward signal, specifically targeting Originality.ai's detection algorithms.
Key Capabilities
- Detector Evasion: Trained to evade AI text detection systems, demonstrating strong generalization across multiple detectors.
- Meaning Preservation: Maintains high semantic similarity (over 0.94) with the original input text.
- Natural Output: Generates fluent and coherent text that reads naturally, suitable for various writing styles.
- Broad Applicability: Effective across diverse domains including academic, technical, and creative writing.
Performance Highlights
AuthorMist Originality shows exceptional performance in reducing AI text detectability, achieving a mean AUROC of 0.49 and a mean F1-score of 0.09 across six major detection systems. It performs particularly well against Hello SimpleAI (AUROC: 0.07), Sapling (AUROC: 0.13), and Winston.ai (AUROC: 0.35).
Good For
- Research into AI text detection: Provides insights into the limitations of current detection systems.
- Privacy-preserving text generation: Useful for contexts where maintaining author privacy or preventing discrimination against AI-assisted writing is permissible and desired.
- Transforming AI-generated drafts: Ideal for refining AI-generated content to achieve a more human-like tone and style without altering core meaning.