Mahal Systems

Overview › Benchmark catalog

ClassEval

Fits code generators that emit whole modules; borrow class-level versus method-level pass@1 reporting to show where multi-part outputs break.

What it asks the agent to do

  • generate a whole Python class with several interdependent methods from a skeleton
  • keep methods consistent with each other when generated one at a time

What to borrow from it

grading design

Caveats

Only 100 hand-written classes (410 methods); data is CC BY-NC 4.0 (non-commercial); 2023 vintage and likely memorized.

Details

Workloadscoding
Interactionsingle turn
Environmentstatic
End state checkedNo
Toolsnone
Gradingunit tests
Size100
Datasynthetic
Personal datanone
LicenseMIT
Harnessinspect_evals, own repo
Statusfrozen

Links