File size: 10,707 Bytes
c10636a | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 | #!/usr/bin/env python3
"""Freeze a leakage-safe train/validation split for reviewed SFT cases."""
from __future__ import annotations
import argparse
import hashlib
import json
from collections import Counter, defaultdict
from pathlib import Path
from typing import Any, Dict, Iterable, Mapping, Sequence
SCHEMA_VERSION = "ifv-sft-case-split-v1"
def _load_jsonl(path: Path) -> list[Dict[str, Any]]:
rows: list[Dict[str, Any]] = []
for line_number, line in enumerate(
path.read_text(encoding="utf-8").split("\n"), start=1
):
if not line.strip():
continue
value = json.loads(line)
if not isinstance(value, dict):
raise ValueError(f"{path}:{line_number} must be a JSON object")
rows.append(value)
return rows
def _index(rows: Sequence[Mapping[str, Any]], *, name: str) -> Dict[str, Mapping[str, Any]]:
result: Dict[str, Mapping[str, Any]] = {}
for row in rows:
case_id = str(row.get("case_id") or "").strip()
if not case_id:
raise ValueError(f"{name} row lacks case_id")
if case_id in result:
raise ValueError(f"duplicate {name} case_id: {case_id}")
result[case_id] = row
return result
def _label(row: Mapping[str, Any]) -> str:
label = str(row.get("label") or row.get("factual_status") or "").strip().lower()
if label in {"supported", "refuted"}:
return label
verdict = str(
row.get("expected_verdict") or row.get("gold_verdict") or ""
).strip().lower()
mapping = {"real": "supported", "fake": "refuted"}
if verdict in mapping:
return mapping[verdict]
raise ValueError("gold labels must be supported/refuted or real/fake")
class _UnionFind:
def __init__(self, values: Iterable[str]) -> None:
self.parent = {value: value for value in values}
def find(self, value: str) -> str:
parent = self.parent[value]
if parent != value:
self.parent[value] = self.find(parent)
return self.parent[value]
def union(self, left: str, right: str) -> None:
left_root = self.find(left)
right_root = self.find(right)
if left_root != right_root:
self.parent[max(left_root, right_root)] = min(left_root, right_root)
def _group_keys(
runtime: Mapping[str, Any],
gold: Mapping[str, Any],
) -> set[str]:
keys = {f"image:{str(runtime.get('image_sha256') or '').casefold()}"}
claim_atom = gold.get("claim_atom")
if isinstance(claim_atom, Mapping):
subject = str(claim_atom.get("subject") or "").strip().casefold()
event = str(claim_atom.get("event") or "").strip().casefold()
if subject and event:
keys.add(f"subject-event:{subject}|{event}")
evidence = gold.get("certifying_evidence")
if isinstance(evidence, list):
for row in evidence:
if not isinstance(row, Mapping):
continue
for field in ("source_id", "snapshot_sha256"):
value = str(row.get(field) or "").strip().casefold()
if value:
keys.add(f"{field}:{value}")
return keys
def build_split(
runtime_rows: Sequence[Mapping[str, Any]],
gold_rows: Sequence[Mapping[str, Any]],
*,
validation_count: int,
validation_supported: int,
seed: str,
prohibited_case_ids: set[str] | None = None,
prohibited_image_hashes: set[str] | None = None,
) -> tuple[list[Dict[str, Any]], Dict[str, Any]]:
runtime = _index(runtime_rows, name="runtime")
gold = _index(gold_rows, name="gold")
if set(runtime) != set(gold):
raise ValueError("runtime and gold case IDs must match exactly")
if validation_count < 1 or validation_count >= len(runtime):
raise ValueError("validation_count must leave non-empty train and validation")
if validation_supported < 0 or validation_supported > validation_count:
raise ValueError("validation_supported is outside validation_count")
blocked_ids = prohibited_case_ids or set()
blocked_hashes = {value.casefold() for value in (prohibited_image_hashes or set())}
overlap_ids = sorted(set(runtime) & blocked_ids)
overlap_hashes = sorted(
{
str(row.get("image_sha256") or "").casefold()
for row in runtime.values()
}
& blocked_hashes
)
if overlap_ids or overlap_hashes:
raise ValueError(
"training candidates overlap prohibited evaluation data: "
f"case_ids={overlap_ids[:5]}, image_hashes={overlap_hashes[:5]}"
)
labels = {case_id: _label(gold[case_id]) for case_id in gold}
union = _UnionFind(runtime)
members_by_key: Dict[str, list[str]] = defaultdict(list)
for case_id in runtime:
for key in _group_keys(runtime[case_id], gold[case_id]):
members_by_key[key].append(case_id)
for members in members_by_key.values():
for member in members[1:]:
union.union(members[0], member)
components: Dict[str, list[str]] = defaultdict(list)
for case_id in runtime:
components[union.find(case_id)].append(case_id)
groups: list[Dict[str, Any]] = []
for members in components.values():
ordered = sorted(members)
group_id = hashlib.sha256(
json.dumps(ordered, separators=(",", ":")).encode("utf-8")
).hexdigest()[:20]
groups.append(
{
"group_id": group_id,
"members": ordered,
"count": len(ordered),
"supported": sum(labels[item] == "supported" for item in ordered),
}
)
groups.sort(
key=lambda group: hashlib.sha256(
f"{seed}:{group['group_id']}".encode("utf-8")
).hexdigest()
)
target = (validation_count, validation_supported)
choices: Dict[tuple[int, int], tuple[int, ...]] = {(0, 0): ()}
for index, group in enumerate(groups):
next_choices = dict(choices)
for (count, supported), selected in choices.items():
state = (
count + int(group["count"]),
supported + int(group["supported"]),
)
if state[0] > target[0] or state[1] > target[1]:
continue
next_choices.setdefault(state, selected + (index,))
choices = next_choices
if target not in choices:
raise ValueError(
"group constraints cannot satisfy requested validation count/labels"
)
validation_groups = {groups[index]["group_id"] for index in choices[target]}
group_by_case = {
case_id: str(group["group_id"])
for group in groups
for case_id in group["members"]
}
rows = [
{
"schema_version": SCHEMA_VERSION,
"case_id": case_id,
"image_sha256": str(runtime[case_id].get("image_sha256") or ""),
"split_group_id": group_by_case[case_id],
"split": (
"validation"
if group_by_case[case_id] in validation_groups
else "train"
),
}
for case_id in sorted(runtime)
]
split_labels = {
split: Counter(labels[row["case_id"]] for row in rows if row["split"] == split)
for split in ("train", "validation")
}
summary = {
"schema_version": SCHEMA_VERSION,
"seed": seed,
"case_count": len(rows),
"group_count": len(groups),
"split_counts": dict(Counter(row["split"] for row in rows)),
"split_label_counts": {
split: dict(sorted(counts.items()))
for split, counts in split_labels.items()
},
"validation_group_ids": sorted(validation_groups),
}
return rows, summary
def _sha256(path: Path) -> str:
digest = hashlib.sha256()
with path.open("rb") as handle:
for chunk in iter(lambda: handle.read(1024 * 1024), b""):
digest.update(chunk)
return digest.hexdigest()
def _args() -> argparse.Namespace:
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument("--runtime", required=True, type=Path)
parser.add_argument("--gold", required=True, type=Path)
parser.add_argument("--prohibited-runtime", type=Path)
parser.add_argument("--output-dir", required=True, type=Path)
parser.add_argument("--validation-count", type=int, default=8)
parser.add_argument("--validation-supported", type=int, default=2)
parser.add_argument("--seed", default="ifv-qwen35-sft-v1")
return parser.parse_args()
def main() -> int:
args = _args()
output_dir = args.output_dir.expanduser().resolve()
if output_dir.exists() and any(output_dir.iterdir()):
raise FileExistsError(f"output directory must be new or empty: {output_dir}")
output_dir.mkdir(parents=True, exist_ok=True)
runtime_rows = _load_jsonl(args.runtime)
gold_rows = _load_jsonl(args.gold)
prohibited_rows = (
_load_jsonl(args.prohibited_runtime)
if args.prohibited_runtime is not None
else []
)
rows, summary = build_split(
runtime_rows,
gold_rows,
validation_count=args.validation_count,
validation_supported=args.validation_supported,
seed=args.seed,
prohibited_case_ids={str(row.get("case_id") or "") for row in prohibited_rows},
prohibited_image_hashes={
str(row.get("image_sha256") or "") for row in prohibited_rows
},
)
split_path = output_dir / "case_split.jsonl"
split_path.write_text(
"\n".join(json.dumps(row, sort_keys=True) for row in rows) + "\n",
encoding="utf-8",
)
manifest = {
**summary,
"inputs": {
"runtime": {"path": str(args.runtime.resolve()), "sha256": _sha256(args.runtime)},
"gold": {"path": str(args.gold.resolve()), "sha256": _sha256(args.gold)},
"prohibited_runtime": (
{
"path": str(args.prohibited_runtime.resolve()),
"sha256": _sha256(args.prohibited_runtime),
}
if args.prohibited_runtime is not None
else None
),
},
"artifacts": {
"case_split": "case_split.jsonl",
"case_split_sha256": _sha256(split_path),
},
}
(output_dir / "manifest.json").write_text(
json.dumps(manifest, ensure_ascii=False, indent=2) + "\n",
encoding="utf-8",
)
print(json.dumps(manifest, ensure_ascii=False, indent=2))
return 0
if __name__ == "__main__":
raise SystemExit(main())
|