Skip to content
Docs
Esc
↑↓navigate↵open⌘Jpreview
On this page

Unreleased

Changes on main that aren't in a PyPI release yet.

  • OSWorld computer-use tasks: scripts/import_osworld.py imports 16 of OSWorld’s (Apache-2.0) LibreOffice Calc and Writer tasks as computer-use environments, each with a reference episode, and writes OSWorld’s license and a NOTICE next to them. They boot the Daytona snapshot deepgym-osworld-libreoffice (--build-snapshot creates it), since Daytona’s default image has no LibreOffice; metadata.json takes a snapshot. The verifier scores the saved file with the task’s own compare_table rules (ported into the shared _checks.py, standard library only) or its reference text, never a screenshot. Rules the starting file already holds, or holds once an empty sheet is added, are guards, so an untouched file or a blank sheet scores 0; a saved file the readers can’t read (missing, corrupt, not a regular file, too large) scores 0, and reading and scoring run in a time- and memory-limited child whose failure also scores 0, so nothing in the file can crash or stall the verifier. deepgym check asks a computer-use folder for reference_actions.json instead of reference_solution.*, and is inconclusive on one off Daytona.