https://0x1za.github.io/tags/evaluation/Recent content in Evaluation on The Constant FactorHugo -- gohugo.ioen-usMon, 15 Jun 2026 00:00:00 +0000https://0x1za.github.io/blog/cev-score/Mon, 15 Jun 2026 00:00:00 +0000https://0x1za.github.io/blog/cev-score/Your OCR pipeline reports a Character Error Rate of 0.04. Clean. Except the page it ran on has three columns, the model merged two of them, and the “error rate” you’re looking at was computed over text in an order that exists nowhere on the page. The number isn’t slightly wrong — it’s measuring something that has no meaning. This is a short, interactive tour of why, and of a metric that gives you a number you can actually decompose.https://0x1za.github.io/blog/cote-score/Thu, 11 Jun 2026 00:00:00 +0000https://0x1za.github.io/blog/cote-score/Your document-layout model scores an mAP of 0.38. Is that good? Bad? Production-ready? You genuinely can’t tell — and that’s not your fault. The metric itself is the problem. This is a 10-minute tour of why, and of a small library that gives you an answer you can act on. The metrics everyone already uses Train a layout model today and three numbers come out of the box, all inherited from general object detection: