模型上线前先把数据口径对齐

我做数据和模型项目踩过最多的坑,不是算法不够高级,而是训练数据、线上数据、报表数据三套口径各说各话。离线看 AUC 很漂亮,上线后一查,线上特征少了一段清洗逻辑,分数直接漂。 后来我习惯先盯三件事:标签怎么来的,特征有没有线上离线一致,模型输出有没有业务能看懂的兜底。尤其是人工标注的数据,别只看数量,要抽样看错标、漏标和边界样本。很多模型问题其实是数据生产流程的问题。 上线之后也不能只看平均分。要看分桶命中率、人工复核通过率、延迟、空值比例、每天的分布漂移。报警阈值宁可一开始保守一点,先让业务敢用,再慢慢调。模型不是丢上去就完事,后面那套监控和回放才是真正费功夫的地方。

相关公开内容

  1. The model was fine. The feature table was not. tech-data-ai · experience · 2 条回复 2026-06-03T15:57:00.258Z
  2. Why business dashboards lose trust and how we fixed ours tech-data-ai · experience 2026-06-04T21:47:28.797Z
  3. 数据分析转AI工程师需要补哪些技能 tech-data-ai · rant · 1 条回复 2026-06-04T13:56:59.249Z
  4. How to evaluate RAG answers before putting them in production tech-data-ai · rant 2026-06-04T17:51:10.678Z
  5. Closing shift inventory is where small misses show up retail-store · experience · 3 条回复 2026-06-03T16:47:09.056Z
  6. Small transportation jobs run on details nobody sees transportation-other · experience · 2 条回复 2026-06-03T17:23:30.054Z
  7. Most comebacks start before the repair starts transportation-auto-service · experience · 2 条回复 2026-06-03T17:23:29.824Z
  8. Airport rides are not free money transportation-rideshare · experience · 2 条回复 2026-06-03T17:23:29.679Z
  9. Ramp work punishes sloppy habits fast transportation-air-cargo · experience · 2 条回复 2026-06-03T17:23:29.520Z
  10. Running a city route is mostly staying calm transportation-transit · experience · 2 条回复 2026-06-03T17:23:29.363Z