SakanaAI ShinkaEvolve 修复子评估结果校验与去重问题,杜绝结果丢失
fix: validate, dedupe and never drop child evaluation results (#225) Evaluation result integrity fixes from the codebase review. - Validate child results like the seed; store violations as incorrect programs with a diagnostic and keep the evaluator's feedback. - Write metrics.json before correct.json, atomically. - Key stored jobs and the retry queue by a per-job key, so a reused PID cannot drop or evict a result. - Retry results that time out or fail to persist, then record a failed generation instead of dropping it; do not block the monitor on a stuck read. - Record the configured language on program rows.
来源与原文
SakanaAI/ShinkaEvolve · 提交动态
作者:Robert Tjarko Lange
发布时间:(北京时间)
阅读原始来源 · github.com ↗完整链接与内容说明
https://github.com/SakanaAI/ShinkaEvolve/commit/0c340c1517a6be35a7964c52d58bdf22a8754065
本站目前仅有 AI 导读与摘要,未展示原文正文;技术细节、代码和完整结论请查看原始来源。
SakanaAI 开源项目 ShinkaEvolve 提交 PR #225,修复评估结果完整性问题:像种子程序一样校验子评估结果,将违规项作为带诊断信息的错误程序存储并保留评估器反馈;metrics.json 先于 correct.json 原子写入;按每任务键(per-job key)存储任务与重试队列,避免复用 PID 丢失结果;超时或持久化失败的评估会重试并记录失败代次而非丢弃,同时在程序行中记录配置的语言。
来源:SakanaAI/ShinkaEvolve · 提交动态 · github.com