rrivera1849/style-aware-paraphraser-mistral7b
rrivera1849/style-aware-paraphraser-mistral7b is a Mistral-7B-Instruct-v0.3 derivative model developed by rrivera1849, fine-tuned in two stages (SFT and Detector-guided DPO) on the Reddit Million Users Dataset. This model specializes in style-aware paraphrasing of machine-generated text, transforming it into the style of a target human author using 16 reference texts. Its primary differentiator is its ability to evade current machine-text detectors, achieving a maximum AUROC of approximately 0.55, making machine-generated content harder to identify.
Loading preview...
Style-Aware Paraphraser (Mistral-7B + DPO)
This model, rrivera1849/style-aware-paraphraser-mistral7b, is a fine-tuned version of mistralai/Mistral-7B-Instruct-v0.3 designed for adversarial paraphrasing. Its core function is to transform machine-generated text to mimic the writing style of a specific human author, making it significantly harder for machine-text detectors to identify.
Key Capabilities
- Style-Aware Paraphrasing: Rephrases text to match a target author's style, conditioned on 16 author exemplars.
- Detector Evasion: Specifically trained to reduce the detectability of machine-generated text by current detectors, achieving a maximum AUROC of approximately 0.55 (lower is better).
- Iterative Refinement: Designed for a multi-stage inference process involving initial paraphrasing by Mistral-7B and subsequent iterative refinement by this model.
- Social Media Optimization: Best suited for social-media-like text (32–128 tokens), trained on Reddit comments, with demonstrated transferability to Amazon reviews and blogs.
Good For
- Research on Machine-Text Detection Limits: Ideal for stress-testing existing detectors, studying stylistic feature retention, and developing new detection defenses.
- Stylistic Transformation: When the goal is to alter the perceived authorship style of machine-generated content.
Limitations
- Trained primarily on Reddit comments; while it transfers to other domains, performance may vary.
- Requires 16 exemplars from the target author for effective style transfer.
- Detector evasion effectiveness decreases with a larger number of documents (N≥25) per author, as distributions become separable again.