This model is a fine-tuned version of answerdotai/ModernBERT-base specifically optimized for mental health crisis detection. It achieves perfect performance on synthetic crisis detection data with 100% accuracy while maintaining zero false positives.

Key Features

✅ 100% Crisis Detection Rate (Target: ≥95%)
✅ 0% False Positive Rate (Target: <10%)
✅ 8,192 token context window (full ModernBERT capacity)
✅ Real-time inference optimized for production use
✅ Clinical-grade metrics for healthcare applications

Quick Start

python

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch

# Load model and tokenizer
tokenizer = AutoTokenizer.from_pretrained("Akashpaul123/modernbert-crisis-detection")
model = AutoModelForSequenceClassification.from_pretrained(
    "Akashpaul123/modernbert-crisis-detection",
    torch_dtype=torch.bfloat16,
    attn_implementation="eager"  # For compatibility
)

# Example usage
def detect_crisis(text):
    inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=8192)
    
    # Remove token_type_ids (ModernBERT doesn't use them)
    if 'token_type_ids' in inputs:
        del inputs['token_type_ids']
    
    with torch.no_grad():
        outputs = model(**inputs)
        probs = torch.softmax(outputs.logits, dim=-1)
        crisis_prob = probs[0][1].item()
    
    return {
        'crisis_detected': crisis_prob > 0.5,
        'confidence': crisis_prob,
        'classification': 'CRISIS' if crisis_prob > 0.5 else 'SAFE'
    }

# Test examples
examples = [
    "I'm feeling great today and looking forward to the weekend!",
    "I feel hopeless and don't see any point in continuing.",
    "Just had a good conversation with my therapist."
]

for text in examples:
    result = detect_crisis(text)
    print(f"Text: {text}")
    print(f"Result: {result['classification']} (confidence: {result['confidence']:.3f})")
    print("---")

Important Notes

For Production Use

Validation Required: Test on real-world mental health data
Human Oversight: Always include human review for crisis cases
Privacy: Ensure HIPAA compliance for healthcare applications

Limitations

Trained on synthetic data - requires validation on real-world text
Perfect metrics may not generalize to all edge cases

Technical Specifications

Architecture: ModernBERT-base with classification head
Parameters: ~139M (base) + 1.5K (classification)
Context Length: 8,192 tokens
Input Format: Raw text (any length, auto-truncated)
Output: Binary classification (crisis/non-crisis) + confidence

Citation

bibtex

@model{modernbert-crisis-detection,
  author = {Akash Paul},
  title = {ModernBERT Crisis Detection: Fine-tuned Mental Health Crisis Detection},
  year = {2024},
  url = {https://huggingface.co/akashpaul123/modernbert-crisis-detection},
  note = {Fine-tuned from answerdotai/ModernBERT-base}
}

Contact

Author: Akash Paul
Username: Akashpaul123

⚠️ Important: This model is for research and supportive applications only. Always consult mental health professionals for clinical decisions.