source/tools/upload_sacb.py

#!/usr/bin/env python3
"""Publish the corpus to HuggingFace as ilintar/SACB.

Two splits:
  test      the 60 selected tasks -- what the benchmark scores by default
  extended  all 129 validated tasks, including the easy `ledger` tier

The file maps are stored as JSON strings rather than nested structs, because a
struct column would force one schema across tasks that legitimately carry
different files. The loader accepts either.
"""
import argparse, json, sys
from pathlib import Path

HERE = Path(__file__).parent
REPO_ID = "ilintar/SACB"


def to_record(row: dict) -> dict:
    return {
        "task_id": row["task_id"],
        "repo": row["repo"],
        "lang": row["lang"],
        "category": row["category"],
        "difficulty": int(row["difficulty"]),
        "instruction": row["instruction"],
        "files": json.dumps(row["files"]),
        "tests": json.dumps(row["tests"]),
        "gold": json.dumps(row["gold"]),
        "fail_to_pass": list(row["fail_to_pass"]),
        "pass_to_pass": list(row["pass_to_pass"]),
        "n_tests": int(row.get("n_tests", 0)),
    }


def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("--private", action="store_true")
    ap.add_argument("--dry-run", action="store_true")
    args = ap.parse_args()

    selected = [json.loads(l) for l in
                (HERE / "agentic-corpus-60.jsonl").read_text().splitlines() if l.strip()]
    everything = [json.loads(l) for l in
                  (HERE / "agentic-corpus.jsonl").read_text().splitlines() if l.strip()]

    from datasets import Dataset, DatasetDict
    ds = DatasetDict({
        "test": Dataset.from_list([to_record(r) for r in selected]),
        "extended": Dataset.from_list([to_record(r) for r in everything]),
    })
    print(ds)

    if args.dry_run:
        print("dry run, not uploading")
        return 0

    ds.push_to_hub(REPO_ID, private=args.private)
    print(f"pushed to https://huggingface.co/datasets/{REPO_ID}")

    card = (HERE / "SACB_CARD.md")
    if card.is_file():
        from huggingface_hub import HfApi
        HfApi().upload_file(
            path_or_fileobj=str(card), path_in_repo="README.md",
            repo_id=REPO_ID, repo_type="dataset",
            commit_message="Add dataset card",
        )
        print("uploaded dataset card")
    return 0


if __name__ == "__main__":
    sys.exit(main())