Patchbench spec v0: schema, verification contract, sourcing plan
Browse files
README.md
ADDED
|
@@ -0,0 +1,58 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# PatchBench
|
| 2 |
+
|
| 3 |
+
A multi-language benchmark for evaluating whether an LLM can **fix a bug correctly and without introducing new bugs**. Each row presents a real issue report plus a large code context (~500–1500 lines); the model under test produces a patch; a per-row test suite then checks that (a) previously failing tests now pass and (b) the rest of the suite stays green.
|
| 4 |
+
|
| 5 |
+
**Status: v0 spec. Rows are being built.**
|
| 6 |
+
|
| 7 |
+
## Row schema
|
| 8 |
+
|
| 9 |
+
| Field | Type | In `tasks` config | Description |
|
| 10 |
+
|---|---|---|---|
|
| 11 |
+
| `id` | string | ✅ | `patchbench-{lang}-{source}-{nnnn}` |
|
| 12 |
+
| `language` | string | ✅ | `python` \| `javascript` \| `typescript` \| `go` \| `java` \| `rust` |
|
| 13 |
+
| `source` | string | ✅ | `real` (mined bug-fix PR) \| `synthetic` (injected bug) |
|
| 14 |
+
| `origin` | string | ✅ | Source dataset, e.g. `swe-bench-live-multilang`, `swesmith-go` |
|
| 15 |
+
| `repo` | string | ✅ | GitHub repo the code came from |
|
| 16 |
+
| `base_commit` | string | ✅ | Commit the buggy state is taken at |
|
| 17 |
+
| `issue_text` | string | ✅ | Issue/PR report the model must diagnose from |
|
| 18 |
+
| `context_files` | list of `{path, content}` | ✅ | ~500–1500 lines total incl. the buggy file(s), callers, and related modules |
|
| 19 |
+
| `buggy_files` | list of string | ✅ | Paths the fix is expected to touch |
|
| 20 |
+
| `test_files` | list of `{path, content}` | ✅ | Test patch materialized as files |
|
| 21 |
+
| `test_cmds` | list of string | ✅ | Commands that run the tests for this row (language-native) |
|
| 22 |
+
| `env_setup_cmds` | list of string | ❌ (harness only) | Install/build commands needed before `test_cmds` |
|
| 23 |
+
| `fail_to_pass` | list of string | ❌ | Tests that MUST fail pre-fix and pass post-fix |
|
| 24 |
+
| `pass_to_pass` | list of string | ❌ | Tests that MUST stay passing post-fix (the regression check) |
|
| 25 |
+
| `gold_patch` | string | ❌ | Reference fix, for harness/self-eval only |
|
| 26 |
+
|
| 27 |
+
Two configs prevent leakage:
|
| 28 |
+
- **`tasks`** (public eval surface): everything an agent sees — issue, context, tests.
|
| 29 |
+
- **`gold`** (harness-only): `fail_to_pass`, `pass_to_pass`, `gold_patch`, `env_setup_cmds`.
|
| 30 |
+
|
| 31 |
+
## Verification contract (every released row)
|
| 32 |
+
|
| 33 |
+
1. **Pre-fix:** apply `test_files` to the repo at `base_commit` → every `fail_to_pass` test fails; every `pass_to_pass` test passes.
|
| 34 |
+
2. **Post-fix:** apply `gold_patch` → every `fail_to_pass` and `pass_to_pass` test passes.
|
| 35 |
+
3. **Determinism:** each check is run twice; any row with a flaky test is dropped.
|
| 36 |
+
4. **No-leak check:** issue text must not name the exact function/line to change (auto-screen + spot manual review).
|
| 37 |
+
|
| 38 |
+
A model's fix is scored: **resolved** iff all `fail_to_pass` pass AND all `pass_to_pass` still pass. Rows also record per-test results so partial behavior (fix works but broke something) is visible — that is the "introduced a new bug" signal.
|
| 39 |
+
|
| 40 |
+
## Sourcing plan (target ~1,500–2,000 rows)
|
| 41 |
+
|
| 42 |
+
| Language | Real source | Synthetic top-up | Target |
|
| 43 |
+
|---|---|---|---|
|
| 44 |
+
| Python | SWE-Gym + SWE-bench Verified/train | SWE-smith-py | ~550 |
|
| 45 |
+
| Rust | SWE-bench-Live rust + rustbench | SWE-smith-rs | ~300 |
|
| 46 |
+
| Java | SWE-bench-Live java + Multi-SWE-bench java | SWE-smith-java | ~300 |
|
| 47 |
+
| Go | SWE-bench-Live go + Multi-SWE-bench go | SWE-smith-go | ~250 |
|
| 48 |
+
| JS/TS | SWE-bench-Live js+ts + Multi-SWE-bench | SWE-smith-js/ts | ~400 |
|
| 49 |
+
|
| 50 |
+
Context expansion is PatchBench's own contribution: no upstream dataset ships sliced 500–1500-line windows. We materialize the repo at `base_commit`, compute the failing-test dependency closure, and slice the modules inside it, filtering out rows whose closure exceeds the window.
|
| 51 |
+
|
| 52 |
+
## Harness
|
| 53 |
+
|
| 54 |
+
Language-native runner (no Docker required, adapted from the SWE-bench grading loop + microsoft/RepoLaunch per-language build/test recipes): clone at `base_commit` → `env_setup_cmds` → apply `test_files` (+ candidate patch) → run `test_cmds` → parse per-test results → grade against F2P/P2P. Per-repo env caching makes verification tractable.
|
| 55 |
+
|
| 56 |
+
## License audit
|
| 57 |
+
|
| 58 |
+
Upstream dataset licenses: SWE-bench-Live (MIT), SWE-Gym (MIT), SWE-smith (MIT). Multi-SWE-bench / PrimeIntellect rows are CC0/“other” (ByteDance community terms) — those rows are re-verified from original permissively-licensed repos before inclusion, or excluded.
|