METR/hcast-public (GitHub README)

https://github.com/METR/hcast-public

"This repo contains source code for a subset of the tasks used in the HCAST: Human-Calibrated Autonomy Software Tasks paper"; tasks conform to the METR Task Standard. Supports anchor 3 (tasks partly open); the rest of the suite is withheld, and no third-party re-run is documented, so M.1 does not reach 4.