Skip to content

Implement real SWE-Bench Lite verification #23

Description

@anilmurty

swe-bench-lite is an experimental scaffold whose scoring is intentionally disabledscore() raises
NotImplementedError because the earlier version only checked whether a tool (e.g. bash) was called, which
is not real bug-fix verification (tjbench/benchmarks/swe_bench_lite.py, tools in
tjbench/agents/swe_bench_tools.py).

Implement genuine verification: check out the target repo at the task's base commit, apply the model's edit,
and run the FAIL_TO_PASS / PASS_TO_PASS test sets to decide pass/fail. Only remove the disabled gate once
scoring reflects real test outcomes.

Acceptance criteria: real per-task pass/fail from executed tests; sandbox/security notes in
docs/swe-bench-lite.md updated; the scoring-disabled gate removed only when verification is genuine; tests
cover the scoring path. This is a large task — scope a PR plan in the issue thread before implementing.

Metadata

Metadata

Assignees

No one assigned

    Labels

    enhancementNew feature or request

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions