chartreuse-verte/prose-rewriter-1.7b-v2.4
The chartreuse-verte/prose-rewriter-1.7b-v2.4 is a 1.7 billion parameter model based on Qwen3-1.7B-Base, specifically designed for rewriting AI-generated prose to sound more human. It improves sentence length variety, reduces grammatical errors, and minimizes changes in meaning compared to its input. This model excels at transforming choppy, LLM-written text into natural-sounding fictional prose, making it ideal for refining creative writing or narrative content.
Loading preview...
Prose Rewriter 1.7B v2.4
This model, chartreuse-verte/prose-rewriter-1.7b-v2.4, is a specialized 1.7 billion parameter prose rewriter built upon Qwen/Qwen3-1.7B-Base. Its primary function is to take AI-generated prose and re-render it to sound more human and natural, while preserving the original semantic meaning. It achieves this by using a rank-32 LoRA adapter merged at a strength of 1.15.
Key Capabilities
- Human-like Prose Generation: Rewrites result in longer sentences (averaging 15.5 words, up from 12.1 in its own output) and less choppy prose, aligning closer to human writing styles.
- High Fidelity: Significantly reduces instances of changed meanings (25.5% down from 29.3%) and ill-formed sentences (3.4% down from 5.8%) compared to previous versions.
- Reduced AI Artifacts: Minimizes negated narration shapes and truncation, leading to more complete and natural outputs.
- Specialized Training: Trained on a diverse corpus of human prose (literary fiction, writing prompts, fanfiction) with a unique "corrupt forward, train backward" methodology, where human text is the target and LLM-slopified versions are inputs.
- Targeted Optimization: Uses reinforcement learning against a fidelity judge to further refine outputs, rewarding semantic preservation and penalizing grammatical errors, choppiness, and repetition.
Good For
- Refining AI-Generated Fictional Prose: Ideal for making text produced by large language models sound more natural and less robotic in narrative contexts.
- Creative Writing Assistance: Enhancing the flow and style of stories, roleplay dialogues, and other fictional content.
- Improving Readability: Transforming overly simplistic or repetitive AI outputs into engaging, human-quality paragraphs.
Limitations
- Not an Instruct Model: Designed for a single task; it does not respond to general instructions.
- Fictional Prose Only: May not perform well on technical documentation or non-narrative text.
- Paragraph-Level Input: Best suited for single paragraphs; longer inputs should be split.
- English Only: Optimized for English narrative register (third and first person fiction, dialogue).