| """Task ids submitted for training in Five Desks v2: the eight no-hint (level 2) diligence tasks, one per domain, | |
| plus the easier tool-free reasoning task. | |
| All nine had mixed outcomes for Qwen3.8-27B in calibration (reports/calibration-v2: mean 0.19 to 0.69 over 16 to | |
| 24 episodes). | |
| """ | |
| SELECTED = [ | |
| "software-repofix-l2", | |
| "industrial-telemetry-l2", | |
| "science-lab-l2", | |
| "office-scheduling-l2", | |
| "finance-receivables-l2", | |
| "math-problems-l2", | |
| "security-audit-l2", | |
| "media-copyedit-l2", | |
| "math-reasoning-l1", | |
| ] | |