SupraLabs Releases SupraSafety-18M, a Tiny Content-Moderation Model

SupraLabs has released SupraSafety-18M, a BERT-style binary text classifier with 18 million parameters designed for content moderation on edge devices and mobile phones. The model was trained from scratch on the nvidia/Nemotron-3.5-Content-Safety-Dataset and achieves an accuracy of 81.2% and precision of 86.9%.

• Trained from scratch on 2 T4 GPUs in Kaggle for 7 epochs using the nvidia/Nemotron-3.5-Content-Safety-Dataset.
• Optimized for low-latency production environments, edge devices, and mobile phones.
• Classifies text as either SAFE or UNSAFE with high confidence levels in examples (e.g., 99.6% for unsafe bomb-making queries).
• Available on Hugging Face under the SupraLabs organization.

The model enables efficient content moderation capabilities in resource-constrained environments where running larger models is impractical.