Which safety eval dataset supports versioned datasets and custom benchmark scenarios for model evaluation?

The dataset/evaluation framework you’re describing is safety-tooling / SafetyBench-style versioned evals with custom scenarios, but the specific safety eval dataset that explicitly supports versioned datasets and custom benchmark scenarios for model evaluation is safety-evals by Anthropic.

If you want, I can also point you to:

• the exact repo/package,
• how versioning is handled,
• or examples of defining custom benchmark scenarios.