FRED Guard: Efficient Financial Compliance Detection with ModernBERT

Large language models deployed in high-stakes financial applications face dual challenges: ensuring factual grounding while maintaining robust safety guardrails. While existing safety classifiers excel at general content moderation, they miss domain-specific compliance violations critical to financial services, including regulatory breaches and misleading investment advice. We present FRED Guard, a lightweight framework that bridges this gap through targeted financial compliance detection in RAG systems. Our approach leverages a synthetic data pipeline using multiple LLMs to overcome the scarcity of labeled financial compliance data. By orchestrating larger models for diverse violation generation and quality evaluation, we transform FinQA/TAT-QA sources into 8,191 high-quality raw training examples spanning regulatory, fiduciary, and market manipulation violations. A 145M-param ModernBERT with two stage fine-tuning achieves 93.2\% F1 on compliance detection while maintaining 66.7\% F1 on general safety benchmarks (WildGuardTest). FRED Guard delivers this performance with 48x fewer parameters and 28x speed compared to baseline guard models, providing an efficient and deployable path for responsible financial AI.