d0rj's picture
Publish mathematics RL 0.2.0 with five framework adapters
0d8f2b5 verified
Raw History Blame Contribute Delete
7.19 kB
"""Offline checker, bounded source audit and module catalog CLI."""
import argparse
import hashlib
import json
import sys
from collections import Counter
from dataclasses import asdict
from itertools import islice
from pathlib import Path
from . import __version__
from .catalog import MODULES
from .data import DATASET, GRADING_VERSION, REVISION, examples
from .worker import Grader
def main():
parser = argparse.ArgumentParser(description=__doc__)
subcommands = parser.add_subparsers(dest="command", required=True)
subcommands.add_parser("catalog")
check = subcommands.add_parser("grade")
check.add_argument("--module", required=True, choices=MODULES)
check.add_argument("--question", required=True)
check.add_argument("--reference", required=True)
check.add_argument("--answer", required=True)
check.add_argument("--timeout", type=float, default=3)
audit = subcommands.add_parser("self-check")
audit.add_argument("--source")
audit.add_argument("--language", choices=("en", "ru"), default="ru")
audit.add_argument(
"--mode", choices=("train", "interpolate", "extrapolate", "all"), default="interpolate"
)
audit.add_argument("--module", action="append", dest="modules", choices=MODULES, default=[])
audit.add_argument("--revision", default=REVISION)
audit.add_argument(
"--limit",
type=int,
default=100,
help="Maximum checked rows; 0 explicitly means the entire selected source",
)
audit.add_argument("--require-all-modules", action="store_true")
audit.add_argument("--output", type=Path)
export = subcommands.add_parser("export", help="Export a bounded selection for a framework")
export.add_argument("framework", choices=("harbor", "agentenv", "nemo"))
export.add_argument("--output", type=Path, required=True)
export.add_argument("--source")
export.add_argument("--language", choices=("en", "ru"), default="ru")
export.add_argument(
"--mode", choices=("train", "interpolate", "extrapolate", "all"), default="interpolate"
)
export.add_argument("--module", action="append", dest="modules", choices=MODULES, default=[])
export.add_argument("--category", action="append", dest="categories", default=[])
export.add_argument("--limit", type=int, default=100)
export.add_argument("--revision", default=REVISION)
export.add_argument("--seed", type=int, default=42)
export.add_argument("--shuffle-buffer", type=int, default=0)
export.add_argument(
"--agent-id", help="AgentEnv deployed agent store ID (optional for episodes)"
)
export.add_argument("--model", help="AgentEnv model identifier")
args = parser.parse_args()
if args.command == "catalog":
print(
json.dumps(
{"version": __version__, "grading_version": GRADING_VERSION, "modules": MODULES},
indent=2,
)
)
return
if args.command == "grade":
with Grader(args.timeout) as grader:
result = grader.grade(args.module, args.question, args.reference, args.answer)
print(json.dumps(asdict(result), ensure_ascii=False, indent=2))
if result.status == "invalid_reference":
sys.exit(2)
return
if args.command == "export":
import importlib
from .adapters.common import select_tasks
rows = select_tasks(
**{
key: getattr(args, key)
for key in (
"source",
"language",
"mode",
"modules",
"categories",
"limit",
"revision",
"seed",
"shuffle_buffer",
)
}
)
adapter = importlib.import_module(f"mathematics_rl.adapters.{args.framework}")
options = (
{"agent_id": args.agent_id, "model": args.model} if args.framework == "agentenv" else {}
)
result = adapter.export_tasks(rows, args.output, **options)
print(
json.dumps(
{
"framework": args.framework,
"tasks": len(rows),
"output": str(args.output),
"manifest_entries": len(result),
},
ensure_ascii=False,
)
)
return
if args.limit < 0:
parser.error("--limit must be nonnegative")
counts, kinds, statuses = Counter(), Counter(), Counter()
errors = []
modes = ("train", "interpolate", "extrapolate") if args.mode == "all" else (args.mode,)
def selected():
for mode in modes:
yield from examples(
source=args.source,
language=args.language,
mode=mode,
modules=args.modules,
revision=args.revision,
)
rows = islice(selected(), args.limit) if args.limit else selected()
with Grader() as grader:
for example in rows:
result = grader.grade(example.module, example.question, example.answer, example.answer)
counts[example.module] += 1
kinds[result.kind] += 1
statuses[result.status] += 1
if result.reward != 1 and len(errors) < 100:
errors.append(
{
"id": example.id,
"module": example.module,
"question": example.question,
"reference": example.answer,
"result": result.to_dict(),
}
)
missing = sorted(set(MODULES) - counts.keys())
report = {
"package_version": __version__,
"grading_version": GRADING_VERSION,
"dataset": DATASET if not args.source else "local",
"dataset_revision": args.revision if not args.source else None,
"language": args.language,
"mode": args.mode,
"source": args.source,
"source_sha256": hashlib.sha256(Path(args.source).read_bytes()).hexdigest()
if args.source
else None,
"checked": sum(counts.values()),
"coverage": dict(sorted(counts.items())),
"kinds": dict(kinds),
"statuses": dict(statuses),
"missing_modules": missing,
"errors": errors,
"evidence": (
"reference acceptance only; not independent mathematical validation or RL training"
),
}
if args.output:
args.output.parent.mkdir(parents=True, exist_ok=True)
args.output.write_text(
json.dumps(report, ensure_ascii=False, indent=2) + "\n", encoding="utf-8"
)
print(
json.dumps(
{k: report[k] for k in ("checked", "statuses", "kinds", "missing_modules")},
ensure_ascii=False,
indent=2,
)
)
if (
not counts
or any(status != "correct" for status in statuses)
or (args.require_all_modules and missing)
):
sys.exit(1)
if __name__ == "__main__":
main()