lieflat less ai tone v2 expands its evaluation corpus
The author released version 2 of the open-source lieflat less ai tone skill on GitHub, expanding the testing methodology and corpora. The first round used 629 articles, while the second used 270 topic-, fact- and length-matched pairs totaling about 5.2 million Chinese characters. According to the author, the evaluation combined 90 scripted indicators, about 430,000 semantic judgments by Jev, and human review of metaphor appropriateness, identifying Europeanized antithesis, dense parallelism and mismatched metaphors as prominent AI-like traits.
The update provides a more rigorous, openly shared framework for measuring and reducing AI-like writing patterns, although its benchmark findings are self-reported by one author.
我做的去 AI 味 skill，lieflat less ai tone，已经更新第二版到 GitHub啦，这一版在语料和测评上都更严谨，花费的精力和金钱也是成倍增长T T 为了弄清哪些特征值得改，我先用629 篇文章做对照，这也是我第一轮测试的主要语料，大家可以看我之前的帖子。这一轮，我在此基础上做了更准确的语料，用270 对同题、同事实、长度对齐的文章复测，总语料约520 万字。 测量方法上，我加入了更全面的测评方式。除了90项指标的脚本测量，我增加了语义模型测评，需要理解句意的部分，我用 Jev 完成约 43 万次模型语义判断（如果换成其他模型确实很难迅速做完，也非常贵）。最后，我还做了人工评测，主要用于评判比喻是否贴切。 这次测评显著的AI味特征主要有：欧化对举句式、密集并列、比喻贴切度低等。不同模型仍然有各自的写作偏好，例如破折号GPT-5.6 用得偏少，Opus 4.6 用得偏多。 我的目标是做最详细最扎实的开源 AI 味写作测评项目！ 地址：https://github.com/larashero3-dotcom/lieflat-less-ai-tone