README.md

metadata

language:
  - en
  - fr
  - es
  - de
  - it
  - pt
  - nl
license: apache-2.0
tags:
  - tabular
  - multi-modal
  - nutriscore
  - nutrition
pipeline_tag: tabular-classification
datasets:
  - hsilvosa/open-food-facts
metrics:
  - accuracy
  - mae
  - r2

Nutri-Score Multi-Modal Tabular Predictor

This model estimates Nutri-Score grades (A, B, C, D, E) and continuous numerical scores when nutritional values are partially or fully missing.

Dataset

Trained on hsilvosa/open-food-facts.

Model Description

Nutri-Score is a front-of-pack nutritional rating system ranging from A (highest nutritional quality) to E (lowest). When product labels have missing macronutrient values (e.g., missing fiber or sodium content), standard rule-based calculators fail.

This multi-modal ensemble combines:

1. Available continuous macronutrients (energy, fat, saturated fat, sugars, proteins, fiber, sodium, salt).
2. Explicit missingness indicator masks for each nutrient.
3. Derived nutrient ratios (sugars-to-carbs, sat-to-total-fat, energy density).
4. Embedded product name, category, and ingredient list text representations.

Performance Metrics

Metric	Score
Grade Top-1 Accuracy	83.33%
Grade Top-2 Accuracy	93.40%
Numeric Score R²	0.9051 (90.5% Variance Explained)
Mean Absolute Error (MAE)	1.95 points
Median Absolute Error	0.95 points
Weighted F1-Score	0.8353

Python Usage Example

from src.models.nutriscore_predictor import NutriScorePredictor
import pandas as pd

predictor = NutriScorePredictor.from_pretrained("your-username/nutriscore-predictor")

sample_product = pd.DataFrame([{
    "product_name": "Organic Oat Drink",
    "categories": "Plant-based beverages",
    "ingredients_text": "Water, oats (12%), sunflower oil, sea salt.",
    "energy-kcal_100g": 45.0,
    "fat_100g": 1.5,
    "sugars_100g": 4.0,
    "proteins_100g": 0.8,
}])

prediction = predictor.predict(sample_product)[0]
print(f"Predicted Grade: {prediction['predicted_grade']}")
print(f"Predicted Score: {prediction['predicted_score']:.2f}")