用 Jev 构建代码库分类器
开发者分享用 Jev 构建代码库分类器,认为这可能解决智能体生成过度工程化代码的问题,并询问下一步测试方向。
作者运行了一个玩具流水线,将基于文本的评分标准分解为结构化答案,比较 Jev 与 Codex Luna 在 FrontierCode 风格代码品味任务上的评分表现。
ran a toy pipeline comparing jev vs codex luna as graders on FrontierCode style "code taste" tasks: decomposed text-based rubrics into structured answers (e.g. are all changes necessary? do all changes consistently use the abstractions required by the task?) -> run both models in