Sugar record screening: research pilot

An interactive demonstration of saved held-out predictions from a small Open Food Facts pilot. Change the recorded sugar value to explore screening behavior. The browser does not train a model or predict for new ingredient lists. No API key, GPU, server, or paid inference service is used.

One objective

Develop and evaluate a model using ingredient information and food category to identify suspicious nutrient records, starting with sugar. A flag means review is needed; it does not establish a database error or printed-label inaccuracy.

Reproduce

With Python 3.11 or newer, in this directory:

python -m pip install -r requirements.txt
python train_demo.py

Open index.html in a browser. Upload the files in this directory to the root of a Hugging Face Static Space. No build command is required. The original CSV is included to reproduce this run; results.json records its SHA-256 and actual library versions. split_manifest.csv records the exact split.

Method

• Use the saved 353-row completeness-filtered pilot, retaining nonempty ingredient/category/group fields, sugar in [0,100], and products with package quantity unit g. This is a preliminary mass-based filter, not verification of the nutrient basis. Do not generalize to the full 353 rows.
• Keep the lowest-barcode row per exact ingredient hash. This does not remove near-duplicate formulations, related brands, or all product versions.
• Split approximately 60/20/20 into train/calibration/test using group splits with fixed seeds 42 and 43. All models use the same split. No hyperparameter tuning on the test set.
• Category baseline: training category median, with training global median for unseen categories.
• Ingredient model: word TF-IDF unigrams/bigrams (up to 5,000 features), one-hot category, Ridge alpha=10, predictions clipped to [0,100]. Features are fitted on training data only. Bigrams capture local word adjacency, not explicit ingredient rank. Ingredient-order modeling is future work.
• Each model has its own 90% nominal split-conformal interval, using the ceil((n_cal+1)*0.9)-th absolute calibration residual and clipping bounds to [0,100]. Exchangeability is an assumption, not established for this convenience sample; coverage is not a product-specific probability or a verified nutritional range.
• Report held-out MAE, R², empirical interval coverage, interval width and untouched-record flags. Original values are unverified; flag rate is not a false-positive rate.
• Divide held-out sugar values by ten as a controlled perturbation. Exclude unchanged zeros from its denominator. Count flags among changed values; this is a sensitivity demonstration, not real-error recall. Some modified values may themselves be plausible.

Limitations

Small convenience sample, mixed markets and categories, incomplete provenance, uncertain as-sold/prepared basis, unverified source values, and remaining formulation overlap. g describes package quantity, not proof of nutrition normalization. Intervals may be broad and miss decimal errors. No laboratory or package-image adjudication is included. The model has not established real-world error-detection performance. A larger focused dataset and adjudicated records are needed.

This is a new restricted-subset experiment. Its results are not a reproduction of the earlier 353-candidate smoke-test figures.

Data attribution

Contains information from Open Food Facts, made available under the Open Database License (ODbL). The included CSV and derived data are shared under ODbL; retain attribution and applicable share-alike terms. Original extract source: https://huggingface.co/datasets/openfoodfacts/product-database . No package images are bundled in this demo.