cogni-x/nutrition-model-q4-gguf

🍽️ Food Analyzer — GGUF Q4_K_M

A quantized GGUF version of Ateeqq/food-analysis, ready for local inference via llama.cpp.

This model analyzes food images and returns structured nutritional information as a JSON object — including dish name, cooking method, calorie estimates, and macronutrient breakdown.

---

🧩 Model Details

Property	Value
Base Model	Qwen/Qwen3-VL-2B-Instruct
Fine-tune Source	Ateeqq/food-analysis (LoRA)
Fine-tune Method	LoRA (merged & exported)
Quantization	Q4_K_M
Format	GGUF
File Size	~1.05 GB
Original Size (F16)	~3.4 GB
Modality	Image + Text
Output Format	JSON
License	openrail

---

🧠 Capabilities

• 🍕 Food Recognition — Identifies dish name and meal type (homemade, restaurant, etc.)
• 🔥 Calorie Estimation — Predicts total calories per serving
• 🥗 Macronutrient Breakdown — Protein, carbohydrates, and fat in grams
• 🍳 Cooking Method Detection — Boiled, fried, grilled, baked, mixed, etc.
• 📏 Portion Estimation — Approximates ingredient quantities

---

📦 Files

File	Quantization	Size
food-analysis-q4.gguf	Q4_K_M	~1.05 GB

---

🚀 Usage with llama.cpp

1. Download

huggingface-cli download cogni-x/nutrition-model-q4-gguf food-analysis-q4.gguf --local-dir .

2. Run inference

./llama-cli \
  -m food-analysis-q4.gguf \
  --image /path/to/food.jpg \
  -p "As a food-analyzer AI, analyze the image and return a single JSON object containing nutritional information. Respond with JSON only. No extra text." \
  -n 512

---

📤 Example Output

{
  "dish_name": "Vegetable Bowl",
  "food_type": "Homemade food",
  "cooking_method": "boiled and mixed",
  "nutritional_summary": {
    "calories_kcal": 500,
    "protein_g": 20.0,
    "carbohydrate_g": 70.0,
    "fat_g": 15.0
  },
  "portion_size": {
    "quinoa": 200,
    "vegetables": 300,
    "sauce": 50
  }
}

---

⚙️ Quantization Details

This model was quantized using llama.cpp:

# Step 1: Merge LoRA adapter with base model
python convert_hf_to_gguf.py ./food-analysis-merged --outfile food-analysis.gguf

# Step 2: Quantize to Q4_K_M
./llama-quantize food-analysis.gguf food-analysis-q4.gguf Q4_K_M

Q4_K_M uses 4-bit quantization with K-quants for higher accuracy on key weight matrices, providing a good balance between file size (~1 GB) and quality.

---

🔗 Related

• Original LoRA adapter: Ateeqq/food-analysis
• Base model: Qwen/Qwen3-VL-2B-Instruct
• Training dataset: Codatta/MM-Food-100K
• Colab demo: Open in Colab

---

📜 Citation

@misc{food_analyzer_qwen3_vl,
  author = {Muhammad Ateeq},
  title = {Food Analyzer Vision-Language Model},
  year = {2026},
  base_model = {Qwen3-VL-2B-Instruct}
}