File size: 6,136 Bytes
087e4a0 | 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 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 | from __future__ import annotations
import argparse
import json
import random
from pathlib import Path
from typing import Any
def read_records(path: Path) -> tuple[list[dict[str, Any]], str]:
if path.suffix == ".jsonl":
rows = [
json.loads(line)
for line in path.read_text(encoding="utf-8").splitlines()
if line.strip()
]
return [dict(row) for row in rows if isinstance(row, dict)], "jsonl"
raw = json.loads(path.read_text(encoding="utf-8"))
if isinstance(raw, list):
return [dict(row) for row in raw if isinstance(row, dict)], "list"
if isinstance(raw, dict):
for key in ("data", "examples", "instances"):
value = raw.get(key)
if isinstance(value, list):
return [dict(row) for row in value if isinstance(row, dict)], key
raise ValueError(f"Unsupported dataset shape: {path}")
def write_records(path: Path, records: list[dict[str, Any]], shape: str) -> None:
path.parent.mkdir(parents=True, exist_ok=True)
if shape == "jsonl":
path.write_text(
"".join(json.dumps(record, ensure_ascii=False) + "\n" for record in records),
encoding="utf-8",
)
return
payload: Any = records if shape == "list" else {shape: records}
path.write_text(json.dumps(payload, ensure_ascii=False, indent=2) + "\n", encoding="utf-8")
def split_records(
records: list[dict[str, Any]],
*,
val_ratio: float,
min_val: int,
max_val: int | None,
seed: int,
) -> tuple[list[dict[str, Any]], list[dict[str, Any]]]:
if not records:
return [], []
shuffled = list(records)
random.Random(seed).shuffle(shuffled)
val_size = max(min_val, int(round(len(shuffled) * val_ratio)))
if max_val is not None:
val_size = min(val_size, max_val)
val_size = max(1, min(val_size, len(shuffled) - 1)) if len(shuffled) > 1 else 1
return shuffled[val_size:], shuffled[:val_size]
def build_parser() -> argparse.ArgumentParser:
parser = argparse.ArgumentParser(
description="Split each task's train file into deterministic train and validation files."
)
parser.add_argument("--config", type=Path, required=True, help="Task-stream config to read.")
parser.add_argument("--output-config", type=Path, help="Updated task-stream config to write.")
parser.add_argument("--output-dir", type=Path, help="Common output directory for split files.")
parser.add_argument("--val-ratio", type=float, default=0.2)
parser.add_argument("--min-val", type=int, default=1)
parser.add_argument("--max-val", type=int)
parser.add_argument("--seed", type=int, default=42)
parser.add_argument("--overwrite", action="store_true")
parser.add_argument("--dry-run", action="store_true")
return parser
def main() -> int:
args = build_parser().parse_args()
if not 0 < args.val_ratio < 1:
raise ValueError("--val-ratio must be between 0 and 1.")
if args.min_val < 0 or (args.max_val is not None and args.max_val < 1):
raise ValueError("Validation limits must be non-negative and --max-val must be positive.")
config_path = args.config.resolve()
config = json.loads(config_path.read_text(encoding="utf-8"))
config_base = config_path.parent.parent if config_path.parent.name == "configs" else config_path.parent
output_root = args.output_dir.resolve() if args.output_dir else None
changed = False
for task_index, task in enumerate(config.get("tasks", [])):
task_name = str(task.get("task_name", "task"))
splits = task.get("splits", {})
if not isinstance(splits, dict) or not splits.get("train"):
print(f"SKIP {task_name}: no train split")
continue
train_path = _resolve_manifest_path(str(splits["train"]), config_base)
records, shape = read_records(train_path)
train_records, val_records = split_records(
records,
val_ratio=args.val_ratio,
min_val=args.min_val,
max_val=args.max_val,
seed=args.seed + task_index,
)
target_dir = output_root / task_name if output_root else train_path.parent
suffix = ".jsonl" if shape == "jsonl" else ".json"
split_train_path = target_dir / f"train_split{suffix}"
split_val_path = target_dir / f"validation_split{suffix}"
print(
f"{task_name}: {len(records)} -> train={len(train_records)} "
f"validation={len(val_records)} ({split_train_path}, {split_val_path})"
)
if args.dry_run:
continue
if not args.overwrite and (split_train_path.exists() or split_val_path.exists()):
raise FileExistsError(
f"Split file already exists for {task_name}; use --overwrite to replace."
)
write_records(split_train_path, train_records, shape)
write_records(split_val_path, val_records, shape)
splits["train"] = _portable_path(split_train_path, config_base)
splits["validation"] = _portable_path(split_val_path, config_base)
splits["val"] = splits["validation"]
splits.setdefault("anchor", splits["validation"])
changed = True
if args.output_config and changed and not args.dry_run:
args.output_config.parent.mkdir(parents=True, exist_ok=True)
args.output_config.write_text(
json.dumps(config, ensure_ascii=False, indent=2) + "\n", encoding="utf-8"
)
elif changed and not args.dry_run:
print("Split files written. Pass --output-config to write an updated task-stream config.")
return 0
def _resolve_manifest_path(raw_path: str, config_base: Path) -> Path:
path = Path(raw_path).expanduser()
return path.resolve() if path.is_absolute() else (config_base / path).resolve()
def _portable_path(path: Path, config_base: Path) -> str:
try:
return str(path.resolve().relative_to(config_base.resolve()))
except ValueError:
return str(path.resolve())
if __name__ == "__main__":
raise SystemExit(main())
|