guildlm/go-swe-bench-v0

go_swe_bench v0 — real Go bug fixes, verified by the Go toolchain

246 tasks from 79 real Go repositories. Each task is a bug-fix commit whose co-committed test is red on the parent and green on the fix. No LLM anywhere in the build.

Mined on 2026-09-19 from the GuildLM Go mining pipeline by inverting the filter that had thrown the tests away (the pipeline was built for SFT data; a benchmark needs the opposite). Every task was verified twice with go test: green at the commit (≥ 1 test actually ran), red at the parent with the commit's test files copied in. parent_status says how it was red: test_fail (191 tasks, an assertion) or compile_error (55, the new test names a symbol the fix introduces).

Why this exists

Single-function Go benchmarks are saturated: on GuildLM's own 48-task go_dev_bench the union of 13 runs solves 48/48 and the best single model 44/48. Nothing can be learned there any more. Agentic benchmarks are Python-first. This is a bench with room, for Go, at $0, built from public permissively-licensed repos (MIT / Apache-2.0 / BSD / MPL / ISC, ≥ 800 stars, ≤ 40 MB).

Files

• go_swe_bench_v0.jsonl — all 246 verified tasks (32 MB).
• go_swe_bench_v0_eval.jsonl — the registered v0 evaluation set: the most recent task per repo whose before-content is ≤ 30,000 chars (51 tasks: 40 test_fail, 11 compile_error).
• stats.json — miner counters.

Task schema

repo, sha, parent, subject, body          the commit and its message (the "issue")
packages                                  Go packages whose tests are the verifier
src_files, before{path: src}, after{...}  the non-test files the fix changed, before and after
test_files, tests{path: src}              the commit's test files (hidden from the model in v0)
parent_status                             test_fail | compile_error
fail_to_pass                              test names red on the parent and green on the fix
gold_patch                                git diff parent..sha for the source files
size                                      before_chars, patch_lines

Registered v0 evaluation (file-level)

Prompt = subject + body + the BEFORE content of every source file; tests hidden. The model returns each file in full in a ```go block headed // file: <path>. Files are written over a parent worktree, the commit's tests copied in, go test runs on the changed packages; pass@1 iff green. Harness, prereg with predictions committed before the first model call, and every generation: guildlm/guild-code · go/crucible (swe_bench_eval.py, PREREG-go-swe-bench-v0-a-bench-with-room.txt, mine_swe_tasks.py).

To reproduce a score you need the repos at the recorded SHAs: datasets/mining/clone_repos.py --repos repos.swe_v0.jsonl --depth 400 in guild-code, then swe_bench_eval.py --load <generations>.

Baselines (v0 eval set, 51 tasks, file-level, temp 0, seed 0, Ollama Q4_K_M served via 32k-context variants)

model	registered pass@1	repaired extractor	+ 12k-token cap	notes
qwen3-coder:30b (A3B)	11/51 (21.6%)	11/51	12/51	compile_error tasks 0/11; passes by before-size tertile 7 · 4 · 0 (→ 1 with the cap lifted)
qwen2.5-coder:7b	0/51	9/51 (17.6%)	9/51	39/51 outputs open the fence twice (```go / // file: x / ```go); the registered regex read them as empty files

| qwen3-coder:30b (A3B), declaration-level edit form (tools/decledit, K=2 toolchain-fed repair loop) | 15/51 (29.4%) | 16/51 (a repeated-declaration bug in the applier, fixed) | — | round 0 = 15, rounds 1–2 added 0: on 21 of the 36 failures the compiler and the existing tests had nothing to say; strict superset of the file-level passes at 40% of the output |

| qwen3-coder:30b, edit form + self-written test in the loop (K=2, --no-peek; the 30B writes the test, never sees the fix) | 17/51 | — | — | r0 16 · r1 15 · r2 17: the gate (15 tests kept: 2 useful, 13 false alarms, measured on the gold fix before the draw) gained its one useful-on-failure task and broke one right fix for one round | | qwen2.5-coder:7b, same harness | 10/51 | — | — | r0 9 (= its file-level number) · +1 by a false-alarm nudge that landed | | router STRICT: 30B preferred, 7B alternate, gate = the 30B's 15 tests | 17/51 | — | — | oracle (union) 20; 0 switches: on the 3 tasks only the 7B solves the writer produced no test | | 30B writer + deterministic form retry (reject an output with no TestXxx, re-ask once) | — | — | — | form fixed 16/16 in one turn, including all 8 that had been cut off at the token cap; useful tests 2 → 2, kept 15 → 20, false alarms 13 → 18, nocompile 12 → 21 | | router STRICT, same files, new gate (20 tests) | 17/51 | — | — | unchanged, 0 switches. On filebrowser, the one router-blind task the retry gave a gate, that test is red on the parent, the gold fix, the 30B's wrong fix and the 7B's correct fix alike | | 30B writer + one untouched neighbouring _test.go from the same package at the parent commit (36/51 tasks have one) | — | — | — | useful tests 2 → 4 (flat across the three previous arms), toothless 10 → 6, nocompile 21 → 23, kept 20 → 21; the writer stops writing tests that pass on the buggy code and starts reaching for the real API | | router STRICT, same files, exemplar gate (21 tests) | 18/51 | — | — | oracle 20. The campaign's first switch: on nebula the gate is red at the parent, green on the gold, failing on the 30B's fix and green on the 7B's — verified against all four file sets | | 30B writer + compile retry (hand the toolchain's own error lines back once; truncated outputs re-asked at 8000 tokens) | — | — | — | 5/20 non-truncated rows compile, and the rate is stratified by error class: type errors 4/11, undefined-symbol errors 0/7. useful 4 → 5, kept 21 → 25, nocompile 23 → 17 | | router STRICT, same files, compile-retry gate (25 tests) | 18/51 | — | — | unchanged. A compile error is two signals: a type error is information, a name error is the model's ignorance of the package handed back to it | | the commit's own test through the self-test harness (no model; the instrument's ceiling) | — | — | — | old isolation (delete every _test.go in the package): useful 31, nocompile 16 — it deleted the package's test fixtures and rejected the correct answer on 16 tasks. Repaired (keep the package's tests, remove only a file that would redeclare a self-test name): useful 35, kept 38, nocompile 10 | | 30B compile-retry writer, rescored under the repaired isolation (same tests, no model call) | — | — | — | useful 5 → 5, false alarms 20 → 19, kept 25 → 24: one row moved. The harness was wrong and the writer still does not use the fixtures it was denied | | 30B writer, its own system prompt (until 2026-09-23 the writer's calls sent the file-level FIX prompt as the system turn; the user turn asked for a test), bare config | — | — | — | useful 2 → 4, no-test 16 → 0, nocompile 11 → 22: the contradiction cost FORM, and form buys almost nothing | | same, full config (neighbouring test + form retry + compile retry) | — | — | — | useful 5 → 7, false alarms 20 → 18, kept 25. The writer's gap is judgement: four routes to form, none past 7 of 35. Router STRICT 18 (unchanged) |

Denominator. 9 tasks (colima, fx, asdf, caddy, gum, gitleaks, beszel, rqlite, vegeta) have a gold test that compiles only against the fix: it names a symbol the fix introduces, so no test that is red at the parent can exist for them, written by anyone. 3 more yield no test block from the gold file. The answerable maximum, measured by the gold itself, is 38 kept / 35 useful of 51. Every "useful" count in the self-test rows above is out of 35, not 48.

Union of the two file-level rows: 14/51; of the two loop rows: 20/51. Both fail the same way at the top of the size range: no file above ~16k chars passed either model at the registered cap. The gold patch of a passing task is ~48 chars (median); of a failing one ~157. Every number carries its extractor: "registered" is the harness at commit 69238f1, "repaired" collapses the fence stutter (commit 736c8f2), and the two are reported side by side because the repair moved one model by 0 and the other by 9. Full logs: go/crucible/RESULT-go-swe-bench-v0-the-room-is-real-and-the-file-is-the-wall.txt. The algorithm arm (declaration-level edits + a toolchain-fed repair loop) is pre-registered in the same directory.

Caveats, stated up front

• v0 is file-level: the model is told which files change. That is easier than the agentic setting (find the files yourself) and harder than function-level. The agentic setting is the next registered draw.
• Commit messages are the "issue". Some are terse. That is the task as it really occurred.
• Contamination: these repos are public and recent models have likely seen the fixes. Scores are still informative relative to each other and against the file-level ceiling; the archive names the model cutoff next to every number.

Part of GuildLM — small, sharp, open specialists, and an honest log of what they can and cannot do. Built on an M1 Max. Total spend: $0.