Evaluating LLM guardrail configs locally with EvalHub

Summary

This installment evaluates local NeMo Guardrails configurations against large-scale risk datasets using EvalHub, following earlier manual testing. EvalHub provides a unified interface for AI evaluation, with provider integrations, reusable benchmarks, Hugging Face dataset inputs, and both local and Kubernetes deployment options. The benchmark configuration maps dataset prompt and label columns to block and pass labels, while balanced safe and risky examples prevent an always-block guardrail from appearing perfect. On the prompt-injection benchmark, the nuanced regex configuration achieved 51.1% accuracy with 13.7 ms median latency; DeBERTa reached 88.93% accuracy with 45.2 ms median latency. The conclusion favors layered guardrails: targeted regex checks can provide speed, with DeBERTa handling other cases, and EvalHub supplies reproducible measurements of the latency–efficacy trade-off.

Takeaways

1. 1

   Risk-only datasets can make a guardrail that blocks every prompt score 100% accuracy, so evaluations should include both safe and risky prompts.

2. 2

   The nuanced regex guardrail achieved 51.1% accuracy and 13.7 ms median latency, while DeBERTa achieved 88.93% accuracy and 45.2 ms median latency.

3. 3

   The post proposes layering targeted regex checks with the more accurate DeBERTa classifier to balance latency and guardrail efficacy.