MiniLM-L12-H384-uncased_Nvidia-Aegis-AI-Safety
| Property | Value |
|---|---|
| Parameter Count | 33.4M parameters |
| Model Type | Multi-label Text Classification |
| Architecture | MiniLM (BERT-based) |
| Tensor Type | F32 |
| Dataset | Nvidia Aegis AI Content Safety Dataset 1.0 |
What is MiniLM-L12-H384-uncased_Nvidia-Aegis-AI-Safety?
This is a specialized content safety model fine-tuned on the Nvidia Aegis AI Content Safety Dataset. It's designed to detect 14 different categories of potentially harmful content, making it particularly valuable for content moderation and AI safety applications. The model achieves a high accuracy of 95.15% on the test set, with specific optimization for minimizing false negatives in safety-critical scenarios.
Implementation Details
The model is built on Microsoft's MiniLM architecture, fine-tuned using a multi-label classification approach. It processes text inputs and returns predictions across 14 safety categories, including controlled substances, criminal planning, harassment, hate speech, and various forms of harmful content.
- Training set size: 3,099 examples
- Test set size: 359 examples
- Evaluation metrics: Accuracy (95.14%), F1 (0.53), Precision (0.67), Recall (0.44)
- Supports both CPU and GPU inference
Core Capabilities
- Multi-label classification across 14 safety categories
- Optimized for reducing false negatives in safety applications
- Efficient inference with both CPU and GPU support
- Comprehensive content safety coverage including harassment, hate speech, violence, and PII detection
Frequently Asked Questions
Q: What makes this model unique?
This model's primary strength lies in its comprehensive coverage of safety categories and specific optimization for minimizing false negatives in content moderation scenarios. It balances efficient architecture (33.4M parameters) with robust safety detection capabilities.
Q: What are the recommended use cases?
The model is ideal for content moderation systems, online platforms, and AI safety applications where detecting potentially harmful content is crucial. It's particularly suited for applications requiring multi-category content analysis with high accuracy requirements.
