POLLCHECK/Llama3.1-bias-sequence-classifier

from transformers import AutoModelForSequenceClassification, AutoTokenizer
import torch
import pandas as pd

# Load the merged model and tokenizer
model_path='POLLCHECK/Llama3.1-bias-sequence-classifier'
tokenizer = AutoTokenizer.from_pretrained(model_path)
model = AutoModelForSequenceClassification.from_pretrained(model_path)
model.eval()

# Function to classify text and return probabilities
def classify_text(text):
    # Tokenize the input text and convert it to lower case
    inputs = tokenizer(text.lower(), return_tensors="pt", truncation=True, max_length=512)
    inputs = {k: v.to(model.device) for k, v in inputs.items()}  # Ensure inputs are on the correct device

    # Perform inference without gradient calculation
    with torch.no_grad():
        outputs = model(**inputs)
    
    # Extract logits from the model output
    logits = outputs.logits
    
    # Compute probabilities using the softmax function
    probabilities = torch.nn.functional.softmax(logits, dim=1).squeeze().cpu().numpy()
    
    # Get the index of the class with the highest probability
    predicted_class = torch.argmax(logits, dim=1).item()
    
    # Extract the confidence score for the predicted class
    confidence = probabilities[predicted_class]
    
    # Map class indices to class labels
    class_mapping = {0: "Biased", 1: "Unbiased"}
    predicted_label = class_mapping[predicted_class]
    
    return predicted_label, confidence, probabilities

# Load the CSV file
df = pd.read_csv('/h/sraza/news-media-bias-plus/classifiers/LLM/data/clean_data.csv')


texts = df['text_content'].tolist()
labels = df['text_label'].tolist()

# Convert labels to lower case for case-insensitive comparison
labels = [label.lower() for label in labels]

# Classify a few sample texts and display ground truth along with probabilities
for text, ground_truth in zip(texts[:5], labels[:5]):  # Classify first 5 texts as an example
    predicted_label, confidence, probabilities = classify_text(text)
    print(f"Text: {text[:100]}...")  # Print first 100 characters
    print(f"Ground Truth: {ground_truth}")
    print(f"Predicted class: {predicted_label}")
    print(f"Confidence: {confidence:.2f}")
    print(f"Probabilities: {probabilities}")
    print("---")