File size: 5,156 Bytes
66cf27f df6fa5a 66cf27f 5f867e3 66cf27f df6fa5a 66cf27f 5f867e3 66cf27f 5f867e3 df6fa5a 5f867e3 66cf27f df6fa5a 66cf27f df6fa5a 66cf27f 5f867e3 66cf27f | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 | """Build the public Hugging Face dataset: task catalog, seeded sample instances, and source."""
import json
import os
import shutil
import sys
from pathlib import Path
ROOT = Path(__file__).resolve().parents[1]
sys.path.insert(0, str(ROOT))
from workbench.tasks import ALL_TASK_IDS, TASK_IDS, build_task # noqa: E402
DOMAINS = {
"swe": "software engineering", "ind": "industrial and physical systems",
"sci": "natural science", "off": "office and white-collar work",
"fin": "finance and economics", "math": "math and formal reasoning",
"sec": "cybersecurity", "med": "media and content production",
}
SUMMARY = {
"swe-fix-function": "Repair a buggy function from a bank of 8 (18 bug variants); hidden tests decide.",
"swe-implement-spec": "Implement a function from a written spec and examples; hidden tests decide.",
"swe-crash-triage": "Reproduce a crash in a small project; report file, line, exception type and the bad record.",
"ind-vibration-alarm": "Apply a consecutive-reading alarm rule to interleaved pump sensor data.",
"ind-tank-balance": "Simulate a tank mass balance from valve and drain schedules; find overflow and a level.",
"ind-oee": "Compute line OEE with overlapping downtime reports merged.",
"sci-longest-orf": "Find and translate the longest open reading frame on a DNA strand.",
"sci-limiting-reagent": "Stoichiometry: limiting reactant and isolated product mass at a given yield.",
"sci-dilution-series": "Beer-Lambert with blank correction and a serial dilution back to stock concentration.",
"off-meeting-slot": "Earliest common meeting slot across calendars in several time zones.",
"off-expense-audit": "Audit expense lines against a policy with caps, currency conversion and receipts.",
"off-regional-revenue": "Join orders to customers and total a quarter's revenue by region.",
"fin-loan-balance": "Amortization with cent rounding and extra principal payments.",
"fin-npv-ranking": "Rank projects by NPV from uneven cash flow schedules.",
"fin-portfolio-value": "Value a multi-currency portfolio from trades, closing prices and FX.",
"math-grid-paths": "Count monotone lattice paths around obstacles.",
"math-number-theory": "CRT, inclusion-exclusion counting, or digit sums of large powers.",
"math-constrained-path": "Shortest path cost and count of shortest routes avoiding a closed node.",
"sec-ssh-bruteforce": "Find the IP that brute-forced and then logged in, per a windowed rule, among decoys.",
"sec-decode-config": "Deobfuscate a defanged config blob (base64 + rotating XOR) to recover a hostname.",
"sec-firewall-rules": "Evaluate first-match firewall rules with CIDRs and port ranges.",
"med-subtitle-stats": "Count a speaker's subtitle cues and their on-screen time, including two-speaker cues.",
"med-edl-duration": "Program duration of a cuts-only EDL, skipping disabled events, as timecode.",
"med-headline-style": "Apply a title-case style guide with small words, acronyms and hyphens.",
"swe-repo-repair": "Find and fix injected regressions in a vendored open-source library; its own test suite grades.",
"swe-repo-boltons": "Repo Repair on boltons 26.2.0.",
"swe-repo-inflection": "Repo Repair on inflection 0.5.1.",
"swe-repo-more-itertools": "Repo Repair on more-itertools 11.1.0.",
"swe-repo-toolz": "Repo Repair on toolz 1.2.0.",
}
def main(out: Path):
if out.exists():
shutil.rmtree(out)
(out / "data").mkdir(parents=True)
with open(out / "data" / "tasks.jsonl", "w") as f:
for tid in ALL_TASK_IDS:
fam, lv = tid.split("@")
f.write(json.dumps({"task_id": f"{fam}-l{lv}", "family": fam, "level": int(lv),
"domain": DOMAINS[fam.split("-")[0]], "summary": SUMMARY[fam]}) + "\n")
with open(out / "data" / "samples.jsonl", "w") as f:
for tid in ALL_TASK_IDS:
if tid.startswith("swe-repo-"): # whole libraries; see data/repo_catalog instead
continue
for seed in range(3):
t = build_task(tid, seed)
f.write(json.dumps({
"task_id": tid, "seed": seed, "instruction": t.instruction,
"files": [{"name": k, "content": v} for k, v in sorted(t.files.items())],
"reference_answer": t.oracle,
}) + "\n")
catalog = Path(os.environ.get("WB_REPO_LIBS", "/app/repo_libs")) / "catalog"
if catalog.is_dir():
shutil.copytree(catalog, out / "data" / "repo_catalog")
src = out / "source"
for rel in ["workbench", "server", "scripts", "tests", "docs", "repo_repair"]:
shutil.copytree(ROOT / rel, src / rel, ignore=shutil.ignore_patterns("__pycache__"))
for rel in ["openenv.yaml", "requirements.txt", "README.md"]:
shutil.copy(ROOT / rel, src / rel)
shutil.copy(ROOT / "README.md", out / "README.md")
print(f"wrote {len(ALL_TASK_IDS)} task variants and {3 * len(ALL_TASK_IDS)} samples to {out}")
if __name__ == "__main__":
main(Path(sys.argv[1] if len(sys.argv) > 1 else "dataset"))
|