README.md

metadata

license: mit
task_categories:
  - text-generation
language:
  - en
tags:
  - shell
  - command-correction
  - benchmark
  - terminal
pretty_name: mash-bench
size_categories:
  - n<1K

mash-bench

203 real-world shell-command repair scenarios for evaluating command correction models — as of Aug 2026, the first benchmark for this task. Each scenario is a wrong/garbled command as a human actually typed it, the stderr it produced (187/203 have it), and one or more acceptable corrections.

{"input": "cd..", "stderr": "-bash: cd..: command not found", "expected": ["cd .."], "source": "nofaq"}
{"input": "git brnch", "stderr": "git: 'brnch' is not a git command. See 'git --help'.", "expected": ["git branch"], "source": "thefuck"}

Sources

• nofaq (139 scenarios) — the artifact of NoFAQ: Synthesizing Command Repairs from Examples (D'Antoni, Singh, Vaughn; ESEC/FSE 2017), MichaelBVaughn/NoFAQ, MIT. Extracted from Synthesis/RuleRepair/TestExamples.fs.
• thefuck (64 scenarios) — inputs, stderr, and fixes taken verbatim from the rule test files of nvbn/thefuck (MIT), spanning git typos, missing sudo, cd errors, package managers, and more.

Scenarios were deduped across sources; entries with fabricated-looking or machine-state-dependent fixes were dropped; alternate corrections were added only where genuinely equivalent. Full cleaning log in PROVENANCE.md.

Scoring

Two modes matter:

• bare input — the model sees only the typed command. Fair for typo-class scenarios (most of thefuck).
• with stderr — the error text is appended to the input. Required for repair-class scenarios (most of nofaq), whose fixes depend on the error (cd files → mkdir files; cd files only given "no such file or directory").

Report both, per source. Exact match against any entry in expected (whitespace-normalized) is the metric; it is conservative — functionally equivalent alternative fixes count as misses unless listed.

Runner: training/bench.py in the fash repo.