identity-hate-detector

Use Model

from transformers import pipeline, AutoTokenizer, AutoModelForSequenceClassification
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
identity_model = AutoModelForSequenceClassification.from_pretrained("Mridul2003/identity-hate-detector").to(device)
identity_tokenizer = AutoTokenizer.from_pretrained("Mridul2003/identity-hate-detector")
identity_inputs = identity_tokenizer(final_text, return_tensors="pt", padding=True, truncation=True)
    if 'token_type_ids' in identity_inputs:
        del identity_inputs['token_type_ids']
    identity_inputs = {k: v.to(device) for k, v in identity_inputs.items()}
    with torch.no_grad():
        identity_outputs = identity_model(**identity_inputs)
    identity_probs = torch.sigmoid(identity_outputs.logits)
    identity_prob = identity_probs[0][1].item()
    not_identity_prob = identity_probs[0][0].item()

    results["identity_hate_custom"] = identity_prob
    results["not_identity_hate_custom"] = not_identity_prob

Offensive Language Classifier (Fine-Tuned on Custom Dataset)

This repository contains a fine-tuned version of the unitary/toxic-bert model for binary classification of offensive language (labels: Offensive vs Not Offensive). The model has been specifically fine-tuned on a custom dataset due to limitations observed in the base model's performance — particularly with identity_hate related content.

---

🔍 Problem with Base Model (unitary/toxic-bert)

The original unitary/toxic-bert model is trained for multi-label toxicity detection with 6 categories:

• toxic
• severe toxic
• obscene
• threat
• insult
• identity_hate

While it performs reasonably well on generic toxicity, it struggles with edge cases involving identity-based hate speech — often:

• Misclassifying subtle or sarcastic identity attacks
• Underestimating offensive content with identity-specific slurs

---

✅ Why Fine-Tune?

We fine-tuned the model on a custom annotated dataset with two clear labels:

• 0: Not Identity Hate
• 1: Identity Hate

The new model simplifies the task into a binary classification problem, allowing more focused training for real-world moderation scenarios.

---

📊 Dataset Overview

• Total examples: ~4,000+
• Balanced between offensive and non-offensive labels
• Contains high proportions of identity_hate, obscene, insult, and more nuanced samples

---

🧠 Model Details

• Base model: unitary/toxic-bert
• Fine-tuned using: Hugging Face 🤗 Trainer API
• Loss function: CrossEntropyLoss (via num_labels=2)
• Batch size: 8
• Epochs: 3
• Learning rate: 2e-5

---

🔬 Performance (Binary Classification)

---