File size: 12,452 Bytes
715cc5a | 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 | from __future__ import annotations
import argparse
import json
from collections import Counter
from pathlib import Path
from src.data.io_utils import read_jsonl, write_csv, write_json
from src.data.normalize_text import stable_hash
def label_distribution(dataset: str, split: str, rows: list[dict], unit: str) -> list[dict]:
counts = Counter(row.get("label") if row.get("label") is not None else "UNLABELED" for row in rows)
return [
{
"dataset": dataset,
"split": split,
"unit": unit,
"label": label,
"count": count,
"percent": round(count / max(1, len(rows)) * 100, 4),
}
for label, count in sorted(counts.items())
]
def claim_hashes(rows: list[dict]) -> set[str]:
hashes = set()
for row in rows:
metadata = row.get("metadata") or {}
hashes.add(metadata.get("claim_norm_hash") or stable_hash(row.get("claim")))
return hashes
def split_overlap_rows(dataset: str, split_rows: dict[str, list[dict]], unit: str) -> list[dict]:
rows: list[dict] = []
names = list(split_rows)
for i, split_a in enumerate(names):
for split_b in names[i + 1 :]:
hashes_a = claim_hashes(split_rows[split_a])
hashes_b = claim_hashes(split_rows[split_b])
overlap = sorted(hashes_a & hashes_b)
rows.append(
{
"dataset": dataset,
"split_a": split_a,
"split_b": split_b,
"unit": unit,
"overlap_type": "claim_norm_hash",
"overlap_count": len(overlap),
"examples": " | ".join(overlap[:5]),
}
)
return rows
def evidence_hashes(rows: list[dict]) -> set[str]:
hashes = set()
for row in rows:
metadata = row.get("metadata") or {}
if metadata.get("evidence_hash"):
hashes.add(metadata["evidence_hash"])
return hashes
def file_exists(path: Path) -> bool:
return path.exists() and path.stat().st_size > 0
def main() -> None:
parser = argparse.ArgumentParser()
parser.add_argument("--data-root", type=Path, default=Path("data_processed"))
parser.add_argument("--stats-dir", type=Path, default=Path("outputs/stats"))
parser.add_argument("--tables-dir", type=Path, default=Path("outputs/tables"))
args = parser.parse_args()
dataset_stats_rows: list[dict] = []
label_rows: list[dict] = []
overlap_rows: list[dict] = []
vifactcheck = {
split: read_jsonl(args.data_root / "vifactcheck" / f"claims_{split}.jsonl")
for split in ["train", "dev", "test"]
}
dataset_stats_rows.append(
{
"Dataset": "ViFactCheck",
"Language": "Vietnamese",
"Domain": "News",
"Train": len(vifactcheck["train"]),
"Dev": len(vifactcheck["dev"]),
"Test": len(vifactcheck["test"]),
"Evidence source": "Context chunks; gold Evidence diagnostic only",
"Labels": "SUPPORTS/REFUTES/NEI",
"Unit": "claim",
}
)
for split, rows in vifactcheck.items():
label_rows.extend(label_distribution("vifactcheck", split, rows, "claim"))
overlap_rows.extend(split_overlap_rows("vifactcheck", vifactcheck, "claim"))
averitec = {
split: read_jsonl(args.data_root / "averitec" / f"claims_{split}.jsonl")
for split in ["train_inner", "dev_inner", "local_test", "hidden_test"]
}
dataset_stats_rows.append(
{
"Dataset": "AVeriTeC",
"Language": "English",
"Domain": "Web fact-checking",
"Train": len(averitec["train_inner"]),
"Dev": len(averitec["dev_inner"]),
"Test": f"local_test={len(averitec['local_test'])}; hidden_test={len(averitec['hidden_test'])}",
"Evidence source": "QA evidence store; hidden test has no local labels",
"Labels": "SUPPORTS/REFUTES/NEI/CONFLICTING",
"Unit": "claim",
}
)
for split, rows in averitec.items():
label_rows.extend(label_distribution("averitec", split, rows, "claim"))
overlap_rows.extend(split_overlap_rows("averitec", averitec, "claim"))
healthver_pairs = {
split: read_jsonl(args.data_root / "healthver" / f"pairs_{split}.jsonl")
for split in ["train", "dev", "test"]
}
dataset_stats_rows.append(
{
"Dataset": "HealthVer",
"Language": "English",
"Domain": "Health/Biomedical",
"Train": len(healthver_pairs["train"]),
"Dev": len(healthver_pairs["dev"]),
"Test": len(healthver_pairs["test"]),
"Evidence source": "claim-evidence pairs",
"Labels": "SUPPORTS/REFUTES/NEI",
"Unit": "pair",
}
)
for split, rows in healthver_pairs.items():
label_rows.extend(label_distribution("healthver", split, rows, "pair"))
healthver_grouped = {
split: read_jsonl(args.data_root / "healthver" / f"claims_grouped_{split}.jsonl")
for split in ["train", "dev", "test"]
}
overlap_rows.extend(split_overlap_rows("healthver", healthver_grouped, "claim_grouped"))
for split_a, split_b in [("train", "dev"), ("train", "test"), ("dev", "test")]:
hashes_a = evidence_hashes(healthver_pairs[split_a])
hashes_b = evidence_hashes(healthver_pairs[split_b])
overlap = sorted(hashes_a & hashes_b)
overlap_rows.append(
{
"dataset": "healthver",
"split_a": split_a,
"split_b": split_b,
"unit": "pair",
"overlap_type": "evidence_text_hash",
"overlap_count": len(overlap),
"examples": " | ".join(overlap[:5]),
}
)
write_csv(args.stats_dir / "dataset_statistics.csv", dataset_stats_rows)
write_csv(args.stats_dir / "label_distribution.csv", label_rows)
write_csv(args.stats_dir / "split_overlap_report.csv", overlap_rows)
write_csv(args.tables_dir / "T1_dataset_statistics.csv", dataset_stats_rows)
label_mapping_rows = [
{"Dataset": "ViFactCheck", "Raw label": "0", "Canonical label": "SUPPORTS", "Main?": "yes", "Note": "verified local numeric label"},
{"Dataset": "ViFactCheck", "Raw label": "1", "Canonical label": "REFUTES", "Main?": "yes", "Note": "verified local numeric label"},
{"Dataset": "ViFactCheck", "Raw label": "2", "Canonical label": "NEI", "Main?": "yes", "Note": "verified local numeric label"},
{"Dataset": "AVeriTeC", "Raw label": "Supported", "Canonical label": "SUPPORTS", "Main?": "yes", "Note": "4-class"},
{"Dataset": "AVeriTeC", "Raw label": "Refuted", "Canonical label": "REFUTES", "Main?": "yes", "Note": "4-class"},
{"Dataset": "AVeriTeC", "Raw label": "Not Enough Evidence", "Canonical label": "NEI", "Main?": "yes", "Note": "4-class"},
{"Dataset": "AVeriTeC", "Raw label": "Conflicting Evidence/Cherrypicking", "Canonical label": "CONFLICTING", "Main?": "yes", "Note": "not merged in main"},
{"Dataset": "HealthVer", "Raw label": "Supports", "Canonical label": "SUPPORTS", "Main?": "yes", "Note": "pair-level"},
{"Dataset": "HealthVer", "Raw label": "Refutes", "Canonical label": "REFUTES", "Main?": "yes", "Note": "pair-level"},
{"Dataset": "HealthVer", "Raw label": "Neutral", "Canonical label": "NEI", "Main?": "yes", "Note": "pair-level"},
]
write_csv(args.tables_dir / "T2_label_mapping.csv", label_mapping_rows)
protocol_rows = [
{"Protocol": "P1", "Dataset": "ViFactCheck", "Input allowed": "Statement + Context chunks", "Forbidden": "Evidence", "Role": "main"},
{"Protocol": "P2", "Dataset": "ViFactCheck", "Input allowed": "Statement + Context-derived WikiKG", "Forbidden": "Evidence", "Role": "proposed"},
{"Protocol": "P3", "Dataset": "ViFactCheck", "Input allowed": "Statement + Evidence", "Forbidden": "report as main", "Role": "upper-bound"},
{"Protocol": "P4", "Dataset": "AVeriTeC", "Input allowed": "Claim + evidence store", "Forbidden": "hidden test label/questions", "Role": "main"},
{"Protocol": "P5", "Dataset": "AVeriTeC", "Input allowed": "Claim + WikiKG evidence", "Forbidden": "hidden test label", "Role": "proposed"},
{"Protocol": "P6", "Dataset": "HealthVer", "Input allowed": "Claim + evidence pair", "Forbidden": "test tuning", "Role": "main"},
{"Protocol": "P7", "Dataset": "HealthVer", "Input allowed": "Claim + evidence-derived WikiKG", "Forbidden": "test tuning", "Role": "proposed"},
{"Protocol": "P8", "Dataset": "All", "Input allowed": "same evidence without KG/path", "Forbidden": "KG features", "Role": "ablation"},
]
write_csv(args.tables_dir / "T3_protocol_matrix.csv", protocol_rows)
averitec_split_report = json.loads((args.stats_dir / "averitec_split_report.json").read_text(encoding="utf-8"))
averitec_local_overlap = [
row
for row in overlap_rows
if row["dataset"] == "averitec"
and {row["split_a"], row["split_b"]} in [{"train_inner", "local_test"}, {"dev_inner", "local_test"}]
]
leakage_checks = {
"raw_manifest_exists": file_exists(args.stats_dir / "raw_file_manifest.csv"),
"vifactcheck_context_evidence_separated": file_exists(args.data_root / "vifactcheck" / "context_chunks.jsonl")
and file_exists(args.data_root / "vifactcheck" / "gold_evidence.jsonl"),
"averitec_local_test_is_official_dev_labeled": all(row.get("label") not in (None, "UNLABELED") for row in averitec["local_test"]),
"averitec_hidden_test_has_no_local_labels": all(row.get("label") is None for row in averitec["hidden_test"])
and not averitec_split_report["official"]["hidden_test_has_labels"],
"averitec_no_train_or_dev_inner_overlap_with_local_test": all(int(row["overlap_count"]) == 0 for row in averitec_local_overlap),
"healthver_pair_files_exist": all(file_exists(args.data_root / "healthver" / f"pairs_{split}.jsonl") for split in ["train", "dev", "test"]),
"healthver_grouped_claim_files_exist": all(
file_exists(args.data_root / "healthver" / f"claims_grouped_{split}.jsonl") for split in ["train", "dev", "test"]
),
"tables_t1_t2_t3_exist": all(file_exists(args.tables_dir / name) for name in ["T1_dataset_statistics.csv", "T2_label_mapping.csv", "T3_protocol_matrix.csv"]),
}
warnings = {
"vifactcheck_official_split_claim_overlap": [
row for row in overlap_rows if row["dataset"] == "vifactcheck" and int(row["overlap_count"]) > 0
],
"averitec_hidden_prediction_only_overlap": [
row
for row in overlap_rows
if row["dataset"] == "averitec" and "hidden_test" in {row["split_a"], row["split_b"]} and int(row["overlap_count"]) > 0
],
"healthver_official_split_claim_overlap": [
row
for row in overlap_rows
if row["dataset"] == "healthver" and row["overlap_type"] == "claim_norm_hash" and int(row["overlap_count"]) > 0
],
"healthver_official_split_evidence_text_overlap": [
row
for row in overlap_rows
if row["dataset"] == "healthver" and row["overlap_type"] == "evidence_text_hash" and int(row["overlap_count"]) > 0
],
}
leakage_report = {
"stage": "Stage 0-1-2 protocol lock",
"checks": leakage_checks,
"pass": all(leakage_checks.values()),
"warnings": warnings,
"notes": {
"averitec_hidden_test": "Hidden-style local file is prediction-only; no metric is computed without official labels.",
"vifactcheck_evidence": "Gold Evidence is diagnostic/upper-bound only, not main input.",
"healthver_unit": "Main unit is pair-level; grouped claims are analysis/robustness files.",
"healthver_overlap": "The official HealthVer split has substantial repeated evidence text across splits; report this risk and consider evidence-disjoint robustness.",
},
}
write_json(args.stats_dir / "leakage_report.json", leakage_report)
print(f"Wrote validation tables and leakage report. PASS={leakage_report['pass']}")
if __name__ == "__main__":
main()
|