Responsible AI Benchmark

The Responsible AI Benchmark consolidates several of the key benchmarks and guardrails to give a baseline view of how the newest LLMs behave. A common set of tests runs across many models, reporting where they cluster on three dimensions:

• Safety: refusal on unsafe prompts, from RabakBench and our WOG risk taxonomy.
• Robustness: factual grounding and abstention on RAG tasks, from KnowOrNot.
• Fairness: differential treatment across groups.

When to use

• Early model and guardrail comparison.
• A baseline view of safety, robustness, and fairness before tuning.

For launch readiness, combine results with Designing an evaluation plan.