ecoicop-pt-bertimbau

ECOICOP Product Classifier (BERTimbau, Portuguese)

This model classifies short product names (product + brand) into 71 ECOICOP categories related to food and beverage products, in Portuguese.
It is based on fine-tuning BERTimbau with data collected from retailers in Portugal.

---

🧠 Model Details

• Architecture: neuralmind/bert-base-portuguese-cased fine-tuned for classification
• Tokenizer: BERT cased tokenizer for Portuguese
• Labels: 71 ECOICOP food & beverage classes
• Framework: Hugging Face Transformers (PyTorch backend)
• Input: Product + Brand (e.g., "Leite meio-gordo Mimosa")
• Output: ECOICOP label (e.g., "01.1.4.1" → Leite meio-gordo ou magro)

---

📊 Performance

Evaluated using a stratified 80/20 split on ≈98k manually verified product entries.

📌 These results show strong generalization across all ECOICOP categories, including those with few examples.

---

📦 Files

• pytorch_model.bin – Trained BERT weights
• config.json – Classification head config
• tokenizer/ – Tokenizer files
• label_encoder.pkl – Scikit-learn LabelEncoder to map back to ECOICOP

---

💡 Intended Use

This model is designed for:

• National Statistical Institutes automating classification in CPI workflows
• Economic researchers studying consumption behavior
• Retail & e-commerce systems tagging items with statistical classification codes

---

🚫 Limitations

• Only handles Portuguese-language product names +
• Restricted to food and beverage ECOICOP categories (not general-purpose)
• Short-text inputs may misclassify rare or ambiguous brand-product combinations

---

🧪 Example Usage

from transformers import BertTokenizer, BertForSequenceClassification
import torch, joblib

# Load model and tokenizer
model = BertForSequenceClassification.from_pretrained("julisses/ecoicop-pt-bertimbau")
tokenizer = BertTokenizer.from_pretrained("julisses/ecoicop-pt-bertimbau")
label_encoder = joblib.load("label_encoder.pkl")  # load this from your local clone or repo

# Predict
inputs = tokenizer("Iogurte natural Danone", return_tensors="pt", truncation=True)
with torch.no_grad():
    logits = model(**inputs).logits
    pred = torch.argmax(logits, axis=1).item()

ecoicop_label = label_encoder.inverse_transform([pred])[0]
print("Predicted ECOICOP code:", ecoicop_label)