EnergyVenom/nutritrack-200-meal-reference-suite

🥗 NutriTrack: 200-Meal International Reference Benchmark & Generalization Suite

The NutriTrack-200-International-Reference-Suite is a standardized, lab-calibrated evaluation dataset and active-learning test suite for automated dietary assessment and multimodal food recognition systems.

📊 Dataset Summary

• Total Benchmark Meals: 200 reference meals with ground-truth nutritional deconstruction.
• Held-Out Active Learning Test Set: 50 distinct, unseen meals for portion generalization auditing.
• Cuisine Categories ($n=7$):
  1. High-Protein & Fitness Foods (25 meals)
  2. South Asian / Indian Regional (50 meals)
  3. Western & American Staples (35 meals)
  4. Mediterranean & Middle Eastern (25 meals)
  5. East Asian & Southeast Asian (30 meals)
  6. Packaged & Barcode Reference Items (20 meals)
  7. Edge Cases & Complex Shared Plates (15 meals)
• Chemical Attribution: USDA FoodData Central SR Legacy & Indian Food Composition Tables (IFCT 2024).

---

🔬 Benchmark Results

Metric	Result	Notes
Top-1 Food Identification	94.8%	Across 200 meals in 7 cuisine categories
Top-3 Food Identification	98.2%
Calorie MAPE	±1.50%	Deterministic (USDA lookup by food ID)
Protein MAPE	±0.80%	Deterministic (USDA lookup by food ID)
Carbs MAPE	±2.10%	Deterministic (USDA lookup by food ID)
Fat MAPE	±1.90%	Deterministic (USDA lookup by food ID)
Median Inference Speed	480ms	Groq LPU Vision fast-path

Why no confidence intervals on MAPE? The calorie/macro numbers come from a deterministic USDA lookup — once the food is correctly identified, the nutrients are a fixed database value, so there is no stochastic variance to put a CI around. The previous version of this file reported [1.50%, 1.50%] with σ = 0.00%, which was technically what the script computed but is misleading to present as a statistical result. The meaningful metric is Top-1 identification accuracy (94.8%), which does have real variance across the test set.

---

🔒 Cryptographic Verification

• Dataset Canonical SHA-256: e2ae4d0648eec1352a68dd85a9b798dec6f9cde92a95d5c92c80d083f11ffefd
• Auditor Bundle SHA-256: 45bf701ebd200dad54f9e01b7280e3705982d1076bee1fabfa3061af75e3a6da

# Clone and verify
git clone https://github.com/SaiPhaniAnirudh/NutriTrack.git
cd NutriTrack
python benchmark/run_benchmark.py --verify-checksum

📜 Citation

@misc{anirudh2026nutritrack,
  author = {Sai Phani Anirudh},
  title = {NutriTrack: Statistically Robust Multimodal AI Food Intelligence with Chemical RAG and Active Learning},
  year = {2026},
  publisher = {GitHub & Hugging Face},
  url = {https://github.com/SaiPhaniAnirudh/NutriTrack}
}