---
title: "Comment by Matt Hanson on \"Building a Fine-Tuned Nutri-Score Classifier.\" | Hashnode"
description: "Really enjoyed the honesty of this one — especially the data audit before any modeling. I research fast-food menus, so I live in the same problem from the other"
url: "https://hashnode.com/posts/building-a-fine-tuned-nutri-score-classifier/6a69be668184976f1c4345fc/comment/6af32537320cba0b78611d58"
---

[Discussion](https://hashnode.com/posts/building-a-fine-tuned-nutri-score-classifier/6a69be668184976f1c4345fc)

Comment [Building a Fine-Tuned Nutri-Score Classifier.](https://hashnode.com/posts/building-a-fine-tuned-nutri-score-classifier/6a69be668184976f1c4345fc)

[Matt Hanson](https://hashnode.com/@matthanson)

[1d ago](https://hashnode.com/posts/building-a-fine-tuned-nutri-score-classifier/6a69be668184976f1c4345fc/comment/6af32537320cba0b78611d58)

Really enjoyed the honesty of this one — especially the data audit before any modeling. I research fast-food menus, so I live in the same problem from the other side: nutrition panels that look complete until you actually sample the rows, and grades that quietly depend on which formula version was used. Your point about the beverage rules and the India eval gap is the part that stuck with me. A milkshake and a bottle of water are both drinks, but they are not scored anything like each other, and a model that misses that will mislead people. Thanks for publishing the audit, not just the accuracy number.