evaluation-building

You can build an AI safety evaluation from a validated design by using the EquiStamp eval pipeline to construct evaluation instruments, generating a build plan document through guided prompts to gather context.

Yes, the Inspect framework is fully supported for building and executing AI safety evaluations, allowing you to set up Inspect tasks and build comprehensive scoring pipelines for your instruments.

Yes, you need Python and the Inspect AI library for execution, along with Docker for containerized environments to properly run the evaluation building pipeline.

The best way to implement Phase 4 of the EquiStamp eval pipeline is to construct evaluation instruments from validated designs using the provided skill, which simplifies the complexity of setting up scoring pipelines.

Constructing evaluation instruments reduces implementation time by simplifying the process of building AI safety evaluations from validated designs, removing the manual complexity of configuring Inspect tasks.