PatchEval Pilot 20

See the leaderboard: patcheval.dogukanurker.com

Run it with BenchKit: github.com/DogukanUrker/BenchKit

PatchEval is a frozen benchmark for agentic coding. Each task starts from the parent of a real Python bug-fix commit whose regression test landed with the fix. The agent receives the reviewed GitHub issue and parent source; scoring is deterministic from hidden fail-to-pass and regression test exit codes.

This immutable pilot-20 release contains 20 tasks mined from recent commits and validated in four cells:

1. the hidden regression test fails on the parent;
2. the parent's previously passing regression suite stays green;
3. the hidden regression test passes on the original fix; and
4. the regression suite stays green after the original fix.

The release publishes no Docker images. BenchKit builds each task runtime locally from its frozen recipe and removes its task-specific Docker resources after the run.

Contents

• dataset.json: release manifest and task IDs
• tasks.jsonl: prompts, commands, runtime recipes, artifact paths, and hashes
• sources/: deterministic parent-commit source archives without .git
• hidden-tests/: grader patches; these remain outside the agent sandbox
• attestations/: validation records produced before freezing
• SHA256SUMS: checksums for every frozen release artifact
• miner/: the source, configuration, schemas, tests, and documentation used to mine, validate, review, and freeze the release

Run with BenchKit

Install the frozen release with BenchKit, then run the stock Pi harness:

git clone https://github.com/DogukanUrker/BenchKit && cd BenchKit
uv run benchkit datasets install patcheval
uv run benchkit --headless --models MODEL --benchmarks patcheval \
  --harness pi --repair-attempts 1

The agent sandbox contains only the checksummed parent source and a restricted inference route. It has no Git history, network egress, hidden tests, gold patch, host mount, or Docker socket. Hidden grader output is never returned to the model; a repair attempt receives only generic feedback.

Integrity

Verify the release after download:

sha256sum --check SHA256SUMS

dataset.json additionally pins the SHA-256 of tasks.jsonl. The task set of a frozen release never changes, and a later mining window becomes a new version. If review finds a hidden test that grades something the issue never asks for, the fix ships as a numbered revision of the same release (the revision field in dataset.json), and BenchKit pins the exact dataset commit, so a cache from an older revision is fetched again.

Revisions

Revision	Date	Change
1	2026-08-23	Initial pilot-20 release.
2	2026-10-27	pygments-2816: the upstream golden accepted only one exact token layout (the apostrophe alone as Text, the words still lexed as JavaScript) that the issue never asks for, and no model matched it in 27 runs. A reviewed hidden test now checks that the issue's own snippets lex with no Error tokens. The repeated four-cell proof passes against 5,310 tests. All other files are byte-identical.

Results from revision 1 are comparable after regrading pygments-2816: run the revision 2 hidden test on each run's saved final patch. The regression result carries over unchanged, since the regression suite never includes the hidden test.

Reproduce the miner

The miner is intentionally separate from BenchKit. From miner/:

uv sync
uv run patcheval-mine --help
uv run pytest

Discovery, automated validation, blind review, and freezing are separate fail-closed gates. See miner/README.md and miner/docs/review-policy.md for the complete method and publication boundary.

Licensing

The miner package declares Apache-2.0 in miner/pyproject.toml. Source snapshots, issue text, and regression-test patches originate from the named upstream repositories and remain subject to their respective licenses. Consult each task's repository field and upstream project before redistribution or reuse.