Upload README.md with huggingface_hub

README.md

@@ -1,26 +1,71 @@
----
-tags:
-- ml-intern
----
+# Item Weight Predictor
 
-# ZZandro/weight-predictor
+A machine learning model that predicts the weight (in grams) of items based on their text description. Designed for food delivery platforms to estimate weights for:
 
-<!-- ml-intern-provenance -->
-## Generated by ML Intern
+- **Menu items** (restaurant food portions)
+- **Grocery items** (packaged food from supermarkets)
+- **Non-food items** (household goods, electronics, toys, etc.)
 
-This model repository was generated by [ML Intern](https://github.com/huggingface/ml-intern), an agent for machine learning research and development on the Hugging Face Hub.
+## Model Architecture
 
-- Try ML Intern: https://smolagents-ml-intern.hf.space
-- Source code: https://github.com/huggingface/ml-intern
+The model uses **per-item-type Ridge regression** on TF-IDF features of text descriptions:
+
+- Separate models trained for each item category (`menu_item`, `grocery`, `non_food`)
+- Log-transformed target to handle wide weight range (grams to kilograms)
+- TF-IDF with 1-3 gram features for robust text representation
+
+## Performance
+
+| Item Type  | MAE    | MAPE   | Training Samples | Validation Samples |
+|------------|--------|--------|------------------|--------------------|
+| Grocery    | 23.1g  | 4.4%   | 17,246           | 3,044              |
+| Menu Item  | 63.1g  | 136.9% | 6,786            | 1,198              |
+| Non-Food   | 479.9g | 113.6% | 7,152            | 1,263              |
+
+Grocery items perform best because most are standardized to ~100g servings. Menu items and non-food have higher variance but still useful for rough estimates.
 
 ## Usage
 
 ```python
-from transformers import AutoModelForCausalLM, AutoTokenizer
+import joblib
+from huggingface_hub import hf_hub_download
+
+# Download and load model
+model_path = hf_hub_download(repo_id="ZZandro/weight-predictor", filename="unified_predictor.pkl")
+predictor = joblib.load(model_path)
+
+# Predict weight
+text = "[MENU_ITEM] grilled chicken breast | food | meal ingredient | portion"
+weight_g = predictor.predict(text, item_type="menu_item")
+print(f"Predicted weight: {weight_g:.1f}g")
+```
+
+### Supported Item Types
 
-model_id = 'ZZandro/weight-predictor'
-tokenizer = AutoTokenizer.from_pretrained(model_id)
-model = AutoModelForCausalLM.from_pretrained(model_id)
+- `menu_item` - Restaurant menu items (e.g., "cheeseburger", "caesar salad")
+- `grocery` - Packaged grocery products (e.g., "chocolate bar", "cereal box")
+- `non_food` - General retail items (e.g., "laptop", "t-shirt", "toy car")
+
+### Text Format
+
+Include the item type tag at the start:
 ```
+[MENU_ITEM] <item description>
+[GROCERY] <item description>
+[NON_FOOD] <item description>
+```
+
+## Data Sources
+
+The training data combines:
+- Amazon product data with shipping weights (non-food items)
+- NutriBench meal descriptions with per-ingredient gram weights (menu items)
+- USDA Foundation Food data with standard serving sizes (grocery items)
+
+## Dataset
+
+The processed training dataset is available at: [ZZandro/item-weight-dataset](https://huggingface.co/datasets/ZZandro/item-weight-dataset)
+
+## License
 
-For non-causal architectures, replace `AutoModelForCausalLM` with the appropriate `AutoModel` class.
+Apache-2.0