Does this sugar value
need a closer look?
Explore how ingredients and food category can help screen unusual nutrition records. A small pilot experiment, with its results and limitations visible.
Saved held-out predictions · Interactive screening1. Select a product and sugar value
Ingredient statement
2. Results for the selected product
g per 100 g, provisional basis · 90% nominal calibration level
Green band: calibrated interval · Dark line: value being tested
Wide intervals may miss errors. Being inside the interval does not confirm that a value is correct.
3. Overall results across the test products
These saved scores summarize all test products. They stay the same when you select another product, change its sugar value, or switch the prediction method above.
| Prediction method | Average error (MAE) ↓ | Model fit (R²) ↑ | Recorded values inside intervals | Average interval width |
|---|
MAE and interval width are in g per 100 g on the provisional source basis. Coverage measures agreement with unverified recorded values. Synthetic flags are not real-error recall; untouched flags are not a measured false-positive rate.
Method, limitations, and reproducible code
We retain products with package mass unit “g”, nonempty ingredient/category information, and sugar between 0 and 100. Package mass is only a proxy: the nutrition basis still requires review. One row per exact ingredient hash is retained before a fixed train/calibration/test split.
Category medians are learned from training data. The ingredient model uses word TF-IDF unigrams/bigrams and one-hot category with Ridge regression. Ingredient position is not explicitly modeled. Both methods use separate calibration residuals to form preliminary intervals. Test products are used only for evaluation.
This is a small, mixed-market convenience sample. Related formulations may remain, labels may contain errors, and calibration assumptions may not hold. Large intervals, sparse categories, and unseen categories limit screening. The next step is a larger focused dataset and package-image verification.
This new restricted-subset experiment is different from the earlier 353-candidate smoke test. All displayed predictions were computed in Python; changing sugar here only reruns the interval comparison. New ingredient lists are not supported.
Training code · Results JSON · Split manifest · Pilot source CSV · Full method and attribution