The Dataset Viewer has been disabled on this dataset.

scale800, Harbor tasks: training split

750 bug-fixing tasks from the scale800_v1 view of harness-hub/agentic-coding-envs, converted to Harbor tasks that run on Hugging Face sandboxes. 22 repositories: Go (SWE-smith), Python (SWE-smith) and Python (SWE-Gym, real upstream commits).

  • tasks/<name>/: task.toml, instruction.md, tests/, solution/, and an environment/ that rebuilds the image.
  • Each task names a prebuilt public image, ghcr.io/younesbelkada/scale800:<name>, with pi, opencode and hermes preinstalled and network access locked.
  • Graded in the agent's container (environment_mode = "shared"): tests/ is uploaded once the agent is done.

The 97 tasks of 5 other repositories are held out in harness-hub/scale800-heldout-harbor. Private: the tasks carry their tests and reference solutions.

Downloads last month
26

Models trained or fine-tuned on harness-hub/scale800-harbor