Skip to content

Roadmap

Planned, in rough order:

  • Lite subset. One task per family with a one-GPU-hour budget, for development runs.
  • Docker image. An official image with pinned dependencies, and the requirement that leaderboard runs use it.
  • Spot re-runs. Every open-track submission with public weights re-run on a sample of cells, with the outcome reported on the pull request.
  • Embeddings mode. A frozen-probe-only path for the global tasks, listed outside the ranking, for models whose weights cannot be released.
  • Data. The Nymeria preprocessing converter, the TransfHAR release and the regenerated HumanML3D wrist simulation (see the gaps in Datasets).
  • Multi-seed rows for the leaderboard's top entries.