| 1 | #!/usr/bin/env python3 |
| 2 | """Select the published SWE-bench Verified subset. |
| 3 | |
| 4 | Deterministic by construction: no random seed, no hand-picking. Re-running this |
| 5 | against the same dataset revision reproduces subset.json byte for byte, so the |
| 6 | sample cannot be quietly tuned after seeing results. |
| 7 | |
| 8 | Rules: |
| 9 | 1. Repos get slots in proportion to their share of the full 500, largest |
| 10 | remainder first, so the sample keeps the benchmark's real composition |
| 11 | (django is ~46% of SWE-bench Verified and stays ~46% here). |
| 12 | 2. Within a repo, instances are ordered by (difficulty, instance_id) and |
| 13 | picked at evenly spaced indices, which spreads the sample across the |
| 14 | repo's own difficulty mix instead of clustering on easy ones. |
| 15 | |
| 16 | Nothing is excluded. psf/requests instances exercise a test suite that makes |
| 17 | live network calls and can exhaust the grader timeout; such a run is reported |
| 18 | as eval_timeout rather than dropped, because silently removing the instances a |
| 19 | harness handles badly is how a benchmark stops meaning anything. |
| 20 | """ |
| 21 | |
| 22 | import json |
| 23 | import sys |
| 24 | from datasets import load_dataset |
| 25 | |
| 26 | TOTAL = 50 |
| 27 | DIFFICULTY_ORDER = {"<15 min fix": 0, "15 min - 1 hour": 1, "1-4 hours": 2, ">4 hours": 3} |
| 28 | |
| 29 | |
| 30 | def allocate(counts, total): |
| 31 | """Largest-remainder apportionment of `total` slots across repo counts.""" |
| 32 | population = sum(counts.values()) |
| 33 | exact = {repo: n * total / population for repo, n in counts.items()} |
| 34 | floors = {repo: int(v) for repo, v in exact.items()} |
| 35 | remaining = total - sum(floors.values()) |
| 36 | order = sorted(exact, key=lambda r: (-(exact[r] - floors[r]), r)) |
| 37 | for repo in order[:remaining]: |
| 38 | floors[repo] += 1 |
| 39 | return floors |
| 40 | |
| 41 | |
| 42 | def evenly_spaced(items, k): |
| 43 | if k <= 0: |
| 44 | return [] |
| 45 | if k >= len(items): |
| 46 | return items |
| 47 | return [items[(2 * i + 1) * len(items) // (2 * k)] for i in range(k)] |
| 48 | |
| 49 | |
| 50 | def main(): |
| 51 | ds = load_dataset("princeton-nlp/SWE-bench_Verified", split="test") |
| 52 | by_repo = {} |
| 53 | for row in ds: |
| 54 | by_repo.setdefault(row["repo"], []).append(row) |
| 55 | |
| 56 | slots = allocate({repo: len(rows) for repo, rows in by_repo.items()}, TOTAL) |
| 57 | |
| 58 | selected = [] |
| 59 | for repo in sorted(by_repo): |
| 60 | rows = sorted( |
| 61 | by_repo[repo], |
| 62 | key=lambda r: (DIFFICULTY_ORDER.get(r["difficulty"], 9), r["instance_id"]), |
| 63 | ) |
| 64 | selected.extend(evenly_spaced(rows, slots[repo])) |
| 65 | |
| 66 | selected.sort(key=lambda r: r["instance_id"]) |
| 67 | out = [ |
| 68 | { |
| 69 | "instance_id": r["instance_id"], |
| 70 | "repo": r["repo"], |
| 71 | "base_commit": r["base_commit"], |
| 72 | "problem_statement": r["problem_statement"], |
| 73 | "difficulty": r["difficulty"], |
| 74 | } |
| 75 | for r in selected |
| 76 | ] |
| 77 | with open("benchmarks/swebench/subset.json", "w", encoding="utf-8") as f: |
| 78 | json.dump(out, f, indent=2, ensure_ascii=False) |
| 79 | f.write("\n") |
| 80 | |
| 81 | print(f"selected {len(out)} of {len(ds)}", file=sys.stderr) |
| 82 | for repo in sorted(by_repo): |
| 83 | share = 100 * len(by_repo[repo]) / len(ds) |
| 84 | print(f" {slots[repo]:2d} {repo:32s} (full set {share:4.1f}%)", file=sys.stderr) |
| 85 | |
| 86 | |
| 87 | if __name__ == "__main__": |
| 88 | main() |
| 89 |