Model Evaluation and Threat Research (METR)
METR is a research nonprofit that works on assessing whether cutting-edge AI systems could pose catastrophic risks to society.
We build the science of accurately assessing risks, so that humanity is informed before developing transformative AI systems.
Read more about our work here.
Our Software
- Inspect AI Cloud Platform
- Public Task Suite
- RE-Bench Task Suite
- Some of our open-source agents can be found at github.com/poking-agents
- Vivaria (deprecated)
Repositories
Showing 10 of 63 repositories
-
hawk Public
Run Inspect AI evals in the cloud
-
inspect-agents Public
A collection of METR wrappers around Inspect agents and of METR scanners for Inspect Scout. Intended to allow consistent usage and customization.
-
bombe-engine Public
METR fork of CharlieBrej/Bombe: the Bombe puzzle engine, assets removed, with eval-specific modifications.