Most writing tools try to change wording; this project targets the particular patterns that make text sound "AI‑generated" and rewrites drafts to feel uneven, specific and human. That single focus — keep every fact unchanged while altering style and structure — lets you paste an LLM draft and get back a version that reads like a person wrote it, without inventing or adding content.
Key Capabilities
- Purpose-built rewrite model: a 12B Gemma-4–based finetune trained on AI-draft → human-original pairs and preference stages so the output changes style and structure while preserving factual tokens (numbers, dates, units, names, quotes).
- Reproducible prompt and sampling: the repo ships a byte-for-byte prompt_format.json and recommends exact sampling settings (temperature 1.0, top_p 0.95, top_k 0, no stop strings) so rewrites are consistent across runtimes.
- Multi-quantization support and local runtimes: GGUF builds (bf16, Q8_0, Q6_K, Q4_K_M, Q3, 2-bit) trade memory for fidelity; supported runtimes include llama.cpp, MLX, vLLM, transformers and an offline desktop app.
- Evaluation and safety posture: external detector checks on the release showed ~95% rewrites judged human and a strict fact judge found no factual problem in 376/420 English rewrites; the project documents typical single‑word slips and urges proofreading for numbers, dates and names.
Who it's for — and tradeoffs
Great fit if you rewrite LLM drafts (emails, essays, reports, forum answers) and need human tone without changing facts. Use it when you control output review and want an offline workflow (local models, command line, desktop app). Look elsewhere if you require guaranteed zero-formatting changes (some headings or lists can merge), absolute perfect Chinese parity (English is stronger), or you cannot audit modified numbers/dates — every release still recommends proofreading those tokens.
How it changes decisions
Instead of prompting LLMs for a human tone, this separates responsibilities: generate a factual draft with any model, then run humanizer to alter prose and structure. That makes quality checks easier (the tool is designed not to add information) and supports low-latency local use with quantized files when GPU/RAM are constrained.
Notes on provenance and license: fine-tuned from google/gemma-4-12B, released under Apache-2.0, and the project explicitly states that no AI detector was used anywhere in training.
