asolinxavier/nutrition-calorie-predictor-v2

🥗 Nutrition Calorie Predictor

Fine-tuned distilgpt2 model that predicts calories per 100g from a food description.

Built as part of the LLM Engineering Course — Week 7 Capstone Project by @asolinxavier

---

🎯 What This Model Does

Give it a food description → it predicts calories per 100g.

Input  : "Steamed fermented rice and lentil cake, soft and fluffy, low fat."
Output : "58"  (kcal per 100g)

---

🏗️ Model Architecture

Property	Value
Base Model	distilgpt2 (82M parameters)
Fine-tuning Method	QLoRA (Parameter Efficient Fine-Tuning)
Trainable Parameters	294,912 (only 0.36% of total!)
LoRA Rank (r)	16
LoRA Alpha	32
Target Module	c_attn
Task	Causal Language Modeling

---

📊 Training Results

Epoch	Train Loss	Val Loss
1	5.383	0.022
2	0.003	0.014 ← Best
5	0.001	0.031
10	0.000	0.036

Best Val Loss: 0.014 ✅

Loss dropped from 5.38 → 0.003 — the model learned dramatically!

---

🍽️ Dataset

• 85 foods total (Indian + Global)
• 42 Indian foods: Idli, Dosa, Biryani, Paneer Butter Masala, Sambar...
• 43 Global foods: Apple, Chicken Breast, Salmon, Oats...
• Split: Train 57 | Val 7 | Test 8
• Dataset: asolinxavier/nutrition_items

---

🚀 How to Use

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel

BASE_MODEL      = "distilgpt2"
FINETUNED_MODEL = "asolinxavier/nutrition-calorie-predictor-v2"

tokenizer  = AutoTokenizer.from_pretrained(BASE_MODEL)
base_model = AutoModelForCausalLM.from_pretrained(BASE_MODEL)
model      = PeftModel.from_pretrained(base_model, FINETUNED_MODEL)
model.eval()

def predict_calories(food_description: str) -> str:
    prompt = (
        f"How many calories are in 100g of this food?\n\n"
        f"{food_description}\n\n"
        f"Calories per 100g: "
    )
    inputs = tokenizer(prompt, return_tensors="pt")
    with torch.no_grad():
        outputs = model.generate(
            **inputs,
            max_new_tokens=10,
            do_sample=False,
            pad_token_id=tokenizer.eos_token_id,
        )
    decoded = tokenizer.decode(outputs[0], skip_special_tokens=True)
    return decoded.split("Calories per 100g:")[-1].strip().split()[0]

# Test it!
print(predict_calories("Steamed fermented rice and lentil cake, soft and fluffy."))
# Output: "58"

---

🗂️ Project Structure

Week 7/
├── day1_qlora_setup.ipynb      ← QLoRA concepts + environment setup
├── day2_dataset_prep.ipynb     ← Dataset preparation + HuggingFace push
├── day3_4_training.ipynb       ← QLoRA fine-tuning (training loop)
├── day5_evaluation.ipynb       ← Model evaluation + custom food testing
└── nutrition/
    ├── items.py                ← NutritionItem data class
    └── evaluator.py            ← Evaluation metrics + charts

---

🔮 Week 8 — Multi-Agent Meal Planner (Next Steps)

This fine-tuned model will be used as the SpecialistAgent inside a 4-agent meal planning framework:

ScannerAgent   → Fetch recipes from TheMealDB API
NutritionAgent → Estimate calories (this model + Claude AI ensemble)
PlannerAgent   → Build personalised weekly meal plan
MessagingAgent → Send daily plan via Pushover notification

---

👤 Author

Asolin Xavier

• HuggingFace: @asolinxavier
• Project: Diet & Nutrition Planner — LLM Engineering Course Week 7 & 8

---

📄 License

MIT License — free to use and modify.