2026

Nemotron SDG

Synthetic data pipeline for safer Nemotron model tuning

Built a synthetic data generation pipeline to fine-tune Nemotron Nano models for malicious-intent detection across agent messages. The workflow combined structured SDG curation, LoRA supervised fine-tuning, and evaluation loops to improve consistency and decision quality. This delivered a 70% increase in structural accuracy and a 5% increase in reasoning accuracy, making downstream safety classification more reliable.

Repository →HF Intent Model →HF Multimodal Model →HF Malicious Message Model →