DocuBench / scripts /run_all.py
urimer's picture
Initial upload: 50 documents, schemas, hand-verified labels, scorer, baseline results (part 2)
8143f06 verified
Raw
History Blame Contribute Delete
5.6 kB
"""run an extraction engine across the full benchmark with idempotent skips.
Currently supports GPT, Claude, and Gemini:
python3 scripts/run_all.py
python3 scripts/run_all.py --engine gpt --doc-id PSU5pciM
python3 scripts/run_all.py --engine claude
python3 scripts/run_all.py --engine gemini
Existing successful result files are skipped and printed. Failed or malformed result
files are rerun unless --force is used, which reruns everything.
"""
from __future__ import annotations
import argparse
import json
import os
import subprocess
import sys
from pathlib import Path
ENGINE_RUNNERS = {
"claude": "scripts/run_claude.py",
"gemini": "scripts/run_gemini.py",
"gpt": "scripts/run_gpt.py",
}
def load_env_file(path: Path) -> dict[str, str]:
env: dict[str, str] = {}
if not path.exists():
return env
with open(path, encoding="utf-8") as f:
for line in f:
stripped = line.strip()
if not stripped or stripped.startswith("#") or "=" not in stripped:
continue
if stripped.startswith("export "):
stripped = stripped[len("export "):].strip()
name, value = stripped.split("=", 1)
name = name.strip()
value = value.strip().strip("'\"")
if name:
env[name] = value
return env
def load_json(path: Path):
with open(path, encoding="utf-8") as f:
return json.load(f)
def benchmark_ids(root: Path) -> list[str]:
return sorted(path.stem for path in (root / "labels").glob("*.json"))
def document_id_map(root: Path) -> dict[str, Path]:
return {path.stem: path for path in (root / "documents").iterdir() if path.is_file()}
def result_state(path: Path) -> str:
if not path.exists():
return "missing"
try:
payload = load_json(path)
except (OSError, json.JSONDecodeError):
return "malformed"
if not isinstance(payload, dict):
return "malformed"
if payload.get("status") == "ok":
return "success"
if payload.get("status") == "failed" or payload.get("error"):
return "tracked_failure" if "data" in payload else "malformed"
return "success" if "data" in payload else "malformed"
def is_successful_result(path: Path) -> bool:
return result_state(path) == "success"
def selected_doc_ids(all_doc_ids: list[str], requested: list[str] | None, limit: int | None) -> list[str]:
if requested:
requested_set = set(requested)
missing = sorted(requested_set - set(all_doc_ids))
if missing:
raise SystemExit(f"unknown doc_id(s): {', '.join(missing)}")
doc_ids = [doc_id for doc_id in all_doc_ids if doc_id in requested_set]
else:
doc_ids = all_doc_ids
return doc_ids[:limit] if limit is not None else doc_ids
def run_one(root: Path, runner: Path, engine: str, doc_id: str, document_path: Path) -> int:
schema_path = root / "schemas" / f"{doc_id}.json"
output_path = root / "results" / engine / f"{doc_id}.json"
cmd = [sys.executable, str(runner), str(document_path), str(schema_path), str(output_path)]
env = os.environ.copy()
env.update(load_env_file(root / ".env"))
completed = subprocess.run(cmd, cwd=root, env=env, check=False)
return completed.returncode
def build_parser() -> argparse.ArgumentParser:
parser = argparse.ArgumentParser(prog="run_all.py")
parser.add_argument("--root", type=Path, default=Path(__file__).resolve().parent.parent)
parser.add_argument("--engine", default="gpt", choices=sorted(ENGINE_RUNNERS))
parser.add_argument("--doc-id", action="append", help="run only this document id; repeatable")
parser.add_argument("--limit", type=int, help="run at most this many selected documents")
parser.add_argument("--force", action="store_true", help="rerun even if a successful result exists")
return parser
def main() -> int:
args = build_parser().parse_args()
root = args.root.resolve()
code_root = Path(__file__).resolve().parent.parent
runner = code_root / ENGINE_RUNNERS[args.engine]
if not runner.exists():
raise SystemExit(f"missing runner: {runner}")
docs = document_id_map(root)
doc_ids = selected_doc_ids(benchmark_ids(root), args.doc_id, args.limit)
skipped = 0
succeeded = 0
tracked_failed = 0
failed = 0
for doc_id in doc_ids:
output_path = root / "results" / args.engine / f"{doc_id}.json"
if not args.force and is_successful_result(output_path):
print(f"skip {doc_id}: existing successful result at {output_path.relative_to(root)}")
skipped += 1
continue
document_path = docs.get(doc_id)
if document_path is None:
print(f"fail {doc_id}: missing source document")
failed += 1
continue
print(f"run {doc_id}: {document_path.relative_to(root)} -> {output_path.relative_to(root)}")
rc = run_one(root, runner, args.engine, doc_id, document_path)
state = result_state(output_path)
if rc == 0 and state == "success":
succeeded += 1
elif rc == 0 and state == "tracked_failure":
tracked_failed += 1
else:
print(f"fail {doc_id}: runner exit={rc}, result_state={state}")
failed += 1
print(
f"done engine={args.engine} "
f"run={succeeded} skipped={skipped} tracked_failed={tracked_failed} failed={failed}"
)
return 1 if failed else 0
if __name__ == "__main__":
raise SystemExit(main())