| from __future__ import annotations |
|
|
| import argparse |
| import json |
| from pathlib import Path |
|
|
| from .data import DecisionRow, load_rows, validate_corpus |
| from .oracles import fixtures, oracle |
| from .teacher import dry_run_plan |
|
|
|
|
| def main() -> None: |
| parser = argparse.ArgumentParser(description="Rook offline curation and guarded teacher utilities") |
| commands = parser.add_subparsers(dest="command", required=True) |
| from .native.commands import add_commands |
| add_commands(commands) |
| validate = commands.add_parser("validate-data") |
| validate.add_argument("paths", nargs="+", type=Path, help="All partitions together, to check leakage") |
| commands.add_parser("teacher-plan") |
| generate = commands.add_parser("write-fixtures") |
| generate.add_argument("path", type=Path) |
| schema = commands.add_parser("write-schema") |
| schema.add_argument("path", type=Path) |
| for name in ("curate-data", "teacher-dry-run", "export-training", "apply-review", "teacher-run", "teacher-ledger-status", "finalize-pilot", "review-ui"): |
| command = commands.add_parser(name) |
| command.add_argument("--root", type=Path, default=Path.cwd()) |
| if name not in {"curate-data", "teacher-ledger-status"}: |
| command.add_argument("--dataset", type=Path, default=Path("data/rook-phase2a-v1")) |
| if name == "export-training": |
| command.add_argument("--output", type=Path, required=True) |
| if name == "apply-review": |
| command.add_argument("--responses", type=Path, required=True) |
| command.add_argument("--version", required=True) |
| command.add_argument("--family-audits", type=Path) |
| if name in {"teacher-dry-run", "teacher-run"}: |
| command.add_argument("--pilot", type=Path) |
| if name == "finalize-pilot": |
| command.add_argument("--config", type=Path, default=Path("configs/teacher.phase2b-small.json")) |
| command.add_argument("--count", type=int, default=5) |
| command.add_argument("--version", required=True) |
| if name == "review-ui": |
| command.add_argument("--store", type=Path, default=Path("runs/review.sqlite3")) |
| command.add_argument("--port", type=int, default=8765) |
| if name in {"teacher-run", "teacher-ledger-status"}: |
| command.add_argument("--ledger", type=Path, default=Path("runs/teacher.sqlite3")) |
| if name == "teacher-run": |
| command.add_argument("--authorization", type=Path, required=True) |
| args = parser.parse_args() |
| if args.command.startswith("native-"): |
| from .native.commands import run |
| try: |
| result = run(args) |
| except (ValueError, RuntimeError, OSError, ImportError, TimeoutError) as exc: |
| parser.exit(2, f"Native command stopped: {exc}\n") |
| print(json.dumps({"status": result.get("status", "recorded"), "output": str(args.output), |
| "check": result.get("check"), "training_rows": result.get("training_rows"), |
| "cuda_checks": result.get("cuda_checks")}, ensure_ascii=False)) |
| return |
| if args.command == "validate-data": |
| try: |
| rows = load_rows(args.paths) |
| for row in rows: |
| if row.provenance.label_method == "program_oracle" and row.source_id == "rook-original-fixtures": |
| if row.options[row.label].key != oracle(row.provenance.oracle_id, row.state): |
| raise ValueError(f"{row.id}: label disagrees with program oracle") |
| except (ValueError, TypeError) as exc: |
| parser.exit(2, f"Validation failed: {exc}\n") |
| print(json.dumps({"status": "valid", "rows": len(rows), "splits": sorted({r.split for r in rows})})) |
| elif args.command == "teacher-plan": |
| print(json.dumps(dry_run_plan(), indent=2)) |
| elif args.command == "write-fixtures": |
| rows = fixtures() |
| validate_corpus(rows) |
| args.path.parent.mkdir(parents=True, exist_ok=True) |
| args.path.write_text("".join(r.model_dump_json(exclude_none=True) + "\n" for r in rows), encoding="utf-8") |
| elif args.command == "write-schema": |
| value = DecisionRow.model_json_schema() |
| value["$schema"] = "https://json-schema.org/draft/2020-12/schema" |
| value["$comment"] = "Use rook validate-data for label bounds, mappings, provenance and whole-corpus isolation; schema alone cannot enforce these." |
| args.path.write_text(json.dumps(value, ensure_ascii=False, indent=2) + "\n", encoding="utf-8") |
| else: |
| from .curation.build import build |
| from .curation.commands import export_training, load_pilot, run_pilot |
| from .curation.review import apply_batch |
| from .teacher import BudgetExceeded |
| from .teacher_ledger import DurableLedger, OwnershipError |
| root = args.root.resolve() |
| dataset = (root / args.dataset).resolve() if hasattr(args, "dataset") else None |
| try: |
| if args.command == "curate-data": |
| build(root) |
| return |
| if args.command == "teacher-dry-run": |
| _, _, _, result = load_pilot(root, dataset, root / args.pilot if args.pilot else None) |
| elif args.command == "export-training": |
| result = export_training(root, dataset, args.output) |
| elif args.command == "apply-review": |
| result = apply_batch(root, dataset, args.responses, args.version, args.family_audits) |
| elif args.command == "finalize-pilot": |
| from .curation.finalize import finalize_pilot |
| result = finalize_pilot(root, dataset, root / args.config, args.version, args.count) |
| elif args.command == "review-ui": |
| import uvicorn |
| from .review_app import create_review_app |
| uvicorn.run(create_review_app(root, dataset, root / args.store), host="127.0.0.1", port=args.port, workers=1) |
| return |
| elif args.command == "teacher-run": |
| result = run_pilot(root, dataset, args.authorization, root / args.ledger, root / args.pilot if args.pilot else None) |
| else: |
| with DurableLedger(root / args.ledger) as ledger: |
| result = ledger.snapshot() |
| except (ValueError, PermissionError, OwnershipError, BudgetExceeded, OSError) as exc: |
| parser.exit(2, f"Command stopped: {exc}\n") |
| print(json.dumps(result, ensure_ascii=False, indent=2)) |
|
|
