ChisatoY/Qwen2.5-7B-ReWrite
ChisatoY/Qwen2.5-7B-ReWrite is a 7.6 billion parameter Qwen2ForCausalLM model, fine-tuned by ChisatoY, specifically designed to rewrite AI-generated text into a more natural, human-like style. Based on Qwen2.5-7B-Instruct, it excels at humanizing both Chinese and English texts while preserving original meaning and formatting, offering a 32,768 token context length.
Loading preview...
Overview
ChisatoY/Qwen2.5-7B-ReWrite is a specialized language model built upon the Qwen2.5-7B-Instruct architecture. Its primary function is to transform AI-generated text into a more natural, human-written style. This 7.6 billion parameter model was fine-tuned using LoRA (Low-Rank Adaptation) and has its optimal LoRA adapter merged back into the base model, making it directly loadable via Transformers with BF16 Safetensors weights. It supports a substantial context length of 32,768 tokens.
Key Capabilities
- AI Text Humanization: Rewrites both Chinese and English AI-style text to sound more human, maintaining original meaning, facts, language, and formatting.
- Bilingual Support: Trained on diverse Chinese and English datasets, including academic abstracts, formal writing, and general text.
- Performance Improvement: Achieved a significant reduction in test loss (approximately 57.14%) compared to the original Qwen2.5-7B-Instruct on an independent English test set, indicating improved humanization capabilities.
Training Details
The model was fine-tuned using supervised fine-tuning (SFT) on a dataset of 43,124 candidates, with 42,261 for training and 863 for validation. The training data was formatted with a specific instruction to rewrite text in a natural human style. Chinese data primarily came from XiangJinYu/Qwen3.5-9B-Humanize-Dataset, focusing on academic and formal texts. English data was sourced from KNipun/ai-humanizer. Training was conducted on a single NVIDIA RTX A6000 48 GB GPU, completing in about 7 hours and 5 minutes.
Known Limitations
- Chinese Text Style: More proficient with academic and formal Chinese texts; less coverage for colloquial, novel, or social media styles.
- Punctuation: Chinese output may feature half-width commas and periods due to training data characteristics.
- Stylistic Nuances: While it simplifies expressions, it may not entirely eliminate promotional tones or generic argumentative structures.
- Accuracy Risk: Rewriting may introduce subtle changes in wording or factual shifts, necessitating human review for critical content.
- Test Set Scope: The independent test set is small and exclusively English, limiting its representativeness for overall bilingual quality assessment.