A Multi-level Gradient Rewriting Dataset of Chinese Academic Paper Abstracts for AIGC Detection MGRD

MGRD sources human-written abstracts from Chinese core journal papers in CNKI and Wanfang Data (2011–2022), covering computer technology, architectural theory, and Chinese drama.

Based on these original abstracts, five LLMs (glm-4.5-air, glm-4.6v, qwen3-14b, deepseek-R1, and gpt-4o-mini) were used to generate AIGC samples at three levels: light polishing, medium rewriting, and heavy rewriting. Notably, the heavy samples were independently generated using only paper titles and keywords. After rule filtering, hierarchical constraint, anomaly removal, semantic consistency verification, perplexity analysis, and expert blind review, we finalized four core data files (light_paired.csv, medium_paired.csv, heavy_paired.csv, all_paired.csv) and one comprehensive file (complete_paired.csv).