hsilvosa/nova-group-classifier

Ingredient-to-NOVA Ultra-Processed Food Classifier

This model is a high-precision multilingual text classification model fine-tuned to predict the NOVA Group (processing level 1 to 4) directly from uncompressed raw ingredient lists of food products worldwide.

Dataset

Trained on hsilvosa/open-food-facts.

Model Description

The NOVA classification system assigns food products into four categories based on the extent and purpose of industrial processing:

• NOVA 1: Unprocessed or minimally processed foods (e.g. fresh fruits, vegetables, plain meat, grains).

• NOVA 2: Processed culinary ingredients (e.g. oils, butter, sugar, salt).

• NOVA 3: Processed foods (e.g. canned vegetables, cheeses, freshly baked breads).

• NOVA 4: Ultra-processed food and drink products (e.g. industrial snacks, soft drinks, pre-packaged meals with additives).

• Pipeline Tag: text-classification

• Input: Uncompressed raw ingredient strings in any major global language.

• Output: Multi-class probability distribution across NOVA 1, 2, 3, and 4.

Performance Metrics

Metric	Score
Top-1 Accuracy	96.20%
Top-2 Accuracy	99.53%
Weighted F1-Score	0.9608
Macro F1-Score	0.8138

Per-Class Performance breakdown

Class	Precision	Recall	F1-Score
NOVA 1 (Minimally Processed)	0.9407	0.9603	0.9504
NOVA 2 (Culinary Ingredients)	0.5333	0.3265	0.4051
NOVA 3 (Processed Foods)	0.9180	0.9138	0.9159
NOVA 4 (Ultra-Processed Foods)	0.9820	0.9861	0.9840

Python Code Example

from src.models.nova_classifier import NOVAClassifier

classifier = NOVAClassifier.from_pretrained("your-username/nova-group-classifier")

ingredient_text = "Whole milk, sugar, cocoa butter, emulsifier (soy lecithin), natural vanilla flavouring"
result = classifier.predict([ingredient_text])[0]

print(f"Predicted Group: {result['label']}")
print(f"Confidence Score: {result['score']:.4f}")

Training Data

Trained on 30,000+ multilingual records from hsilvosa/open-food-facts, covering global products with verified ingredient text strings and ground-truth NOVA assignments.