Datasets:

---

mercor

/

apex-accounting

like 0

Follow

Mercor 104

What changed

• Correct the rubric field names and criterion text in the public task records.
• Add runner-ready tasks_and_rubrics.json and world_descriptions.json metadata.
• Key task-specific filesystem overlays by stable task ID.
• Package the initial world snapshot for evaluation runners.
• Place the QuickBooks exports under the runner's .apps_data convention.
• Update the dataset card and Git LFS rules to describe and track the corrected layout.

Why

The published dev-set snapshot was browsable on the Hub but did not match the layout and metadata contracts expected by the Archipelago evaluation runner. That prevented a clean end-to-end load without local normalization.

Impact

The 10 public accounting tasks can now be loaded directly by the runner with their grading criteria, shared world data, and task-specific overlays resolved consistently.

Validation

• Validated all 10 tasks and 89 rubric criteria.
• Confirmed all 16 task-specific files resolve from their task-ID overlays.
• Confirmed all 90 shared world files and every context_files reference resolve.
• Verified the packaged world ZIP against the browsable world tree.
• Ran Git LFS integrity checks successfully.
• Completed one end-to-end Archipelago smoke task, including environment population, agent execution, and grading.