Mahal Systems

Overview › Benchmark catalog

SWE-smith

Fits teams that need many cheap coding cases from their own repository: point it at a repo, it breaks tests and writes issues. Borrow the bug-injection and keep-only-if-a-test-fails filter.

What it asks the agent to do

  • fix a synthetically injected bug in a real repo from a generated issue text
  • localize the file that a planted failure comes from

What to borrow from it

case templateprivacy patternsandbox design

Caveats

Bugs and issue texts are synthetic, so they are easier and more uniform than real issues; a training toolkit, not a comparable leaderboard.

Details

Workloadscoding
Interactionautonomous long horizon
Environmentcontainers
End state checkedYes
Toolsshell, file editor
Gradingunit tests
Size52,000
Datamixed
Personal datapossible
LicenseMIT
Harnessown repo
Statusactive

Links