RoBERTa for Multi-label Classification of Policy Instruments

This model fine-tunes roberta-base for multilabel classification of policies, targets, and themes.

Model Details

Base model: roberta-base
Max length: 512
Output: 67 multilabel classes (PI - Policy Instrument, TG - Target Group, TH - Theme). There are three main classes that have further sub-categories in them.
Threshold: 0.25

Intended Use

Classify policy documents descriptions into thematic categories.

How to Use

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch
import numpy as np
import joblib
import requests

model_path = "toqeerehsan/multilabel-indicator-classification"
model = AutoModelForSequenceClassification.from_pretrained(model_path)
tokenizer = AutoTokenizer.from_pretrained(model_path)

mlb_url = "https://huggingface.co/toqeerehsan/multilabel-indicator-classification/resolve/main/mlb.pkl"
mlb_path = "mlb.pkl"

with open(mlb_path, "wb") as f:
    f.write(requests.get(mlb_url).content)
mlb = joblib.load(mlb_path)

text = "This program supports clean technology and sustainable development in industries."

inputs = tokenizer(text, return_tensors="pt", truncation=True, padding=True, max_length=512)

model.eval()
with torch.no_grad():
    logits = model(**inputs).logits
    probs = torch.sigmoid(logits).squeeze().numpy()

# Threshold
binary_preds = (probs > 0.25).astype(int)
predicted_labels = [label for i, label in enumerate(mlb.classes_) if binary_preds[i] == 1]

print("Predicted Labels:", predicted_labels)

# Predicted Labels: ['PI007', 'PI008', 'TG20', 'TG21', 'TG22', 'TG25', 'TG29', 'TG31', 'TH31']