File size: 10,467 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 | from __future__ import annotations
import argparse
from pathlib import Path
from typing import Any
from src.data.io_utils import read_jsonl, write_csv, write_jsonl
from src.data.normalize_text import normalize_whitespace
DATASET_SPECS: dict[str, dict[str, Any]] = {
"healthver": {
"protocol": "P6_pair_verification",
"id_field": "pair_id",
"splits": {
"train": "data_processed/healthver/pairs_train.jsonl",
"dev": "data_processed/healthver/pairs_dev.jsonl",
"test": "data_processed/healthver/pairs_test.jsonl",
},
"candidate_template": "outputs/retrieval/healthver/candidate_pool_{split}.jsonl",
},
"vifactcheck": {
"protocol": "P1_full_context_retrieval",
"id_field": "claim_id",
"splits": {
"train": "data_processed/vifactcheck/claims_train.jsonl",
"dev": "data_processed/vifactcheck/claims_dev.jsonl",
"test": "data_processed/vifactcheck/claims_test.jsonl",
},
"candidate_template": "outputs/retrieval/vifactcheck/candidate_pool_{split}.jsonl",
},
"averitec": {
"protocol": "P4_open_retrieval",
"id_field": "claim_id",
"qa_evidence": "data_processed/averitec/qa_evidence.jsonl",
"splits": {
"train_inner": "data_processed/averitec/claims_train_inner.jsonl",
"dev_inner": "data_processed/averitec/claims_dev_inner.jsonl",
"local_test": "data_processed/averitec/claims_local_test.jsonl",
},
"candidate_template": "outputs/retrieval/averitec/candidate_pool_{split}.jsonl",
},
}
def candidate_summary(candidate: dict[str, Any]) -> dict[str, Any]:
metadata = candidate.get("metadata") if isinstance(candidate.get("metadata"), dict) else {}
return {
"candidate_id": candidate.get("candidate_id"),
"text": candidate.get("text", ""),
"final_rank": candidate.get("final_rank"),
"reranker_score": candidate.get("reranker_score"),
"is_gold": candidate.get("is_gold"),
"source_type": candidate.get("source_type"),
"question": candidate.get("question") or metadata.get("question"),
"answer": candidate.get("answer") or metadata.get("answer"),
}
def select_evidence(dataset: str, candidates: list[dict[str, Any]], top_k: int) -> list[dict[str, Any]]:
selected = candidates[:top_k]
if dataset != "healthver":
return selected
anchors = [row for row in candidates if row.get("source_type") == "paired_evidence"]
if not anchors:
return selected
anchor = anchors[0]
selected_ids = {anchor.get("candidate_id")}
augmentations = [row for row in candidates if row.get("candidate_id") not in selected_ids]
return [anchor] + augmentations[: max(0, top_k - 1)]
def format_input_text(dataset: str, claim: str, evidence: list[dict[str, Any]], input_format: str = "flat") -> str:
claim = normalize_whitespace(claim)
if dataset == "vifactcheck":
parts = [f"[STATEMENT] {claim}"]
for idx, item in enumerate(evidence, start=1):
parts.append(f"[CONTEXT_CHUNK_{idx}] {normalize_whitespace(item.get('text', ''))}")
return "\n".join(parts)
if dataset == "healthver":
parts = [f"[CLAIM] {claim}"]
aug_idx = 1
for item in evidence:
text = normalize_whitespace(item.get("text", ""))
if item.get("source_type") == "paired_evidence":
parts.append(f"[EVIDENCE] {text}")
else:
parts.append(f"[AUGMENTED_EVIDENCE_{aug_idx}] {text}")
aug_idx += 1
return "\n".join(parts)
parts = [f"[CLAIM] {claim}"]
if dataset == "averitec" and input_format == "qa":
for idx, item in enumerate(evidence, start=1):
question = normalize_whitespace(item.get("question", ""))
answer = normalize_whitespace(item.get("answer", ""))
text = normalize_whitespace(item.get("text", ""))
if question:
parts.append(f"[QUESTION_{idx}] {question}")
if answer:
parts.append(f"[ANSWER_{idx}] {answer}")
parts.append(f"[EVIDENCE_{idx}] {text}")
return "\n".join(parts)
for idx, item in enumerate(evidence, start=1):
parts.append(f"[EVIDENCE_{idx}] {normalize_whitespace(item.get('text', ''))}")
return "\n".join(parts)
def load_candidate_rows(path: Path) -> dict[str, dict[str, Any]]:
rows: dict[str, dict[str, Any]] = {}
for row in read_jsonl(path):
rows[row["query_id"]] = row
return rows
def load_averitec_qa_by_evidence(path: Path) -> dict[str, dict[str, Any]]:
rows: dict[str, dict[str, Any]] = {}
if not path.exists():
return rows
for row in read_jsonl(path):
metadata = row.get("metadata") if isinstance(row.get("metadata"), dict) else {}
evidence_id = metadata.get("evidence_id")
if evidence_id:
rows[str(evidence_id)] = row
return rows
def enrich_averitec_qa(candidates: list[dict[str, Any]], qa_by_evidence: dict[str, dict[str, Any]]) -> list[dict[str, Any]]:
if not qa_by_evidence:
return candidates
enriched: list[dict[str, Any]] = []
for candidate in candidates:
updated = dict(candidate)
metadata = dict(updated.get("metadata") if isinstance(updated.get("metadata"), dict) else {})
qa_row = qa_by_evidence.get(str(updated.get("candidate_id")))
if qa_row:
updated["question"] = qa_row.get("question")
updated["answer"] = qa_row.get("answer")
metadata.setdefault("question_id", qa_row.get("question_id"))
metadata.setdefault("answer_type", (qa_row.get("metadata") or {}).get("answer_type"))
metadata.setdefault("source_url", qa_row.get("source_url"))
updated["metadata"] = metadata
enriched.append(updated)
return enriched
def build_dataset_split(
dataset: str,
split: str,
source_path: Path,
candidate_path: Path,
output_path: Path,
top_k: int,
protocol: str,
id_field: str,
input_format: str,
qa_by_evidence: dict[str, dict[str, Any]] | None = None,
) -> dict[str, Any]:
source_rows = read_jsonl(source_path)
candidates_by_id = load_candidate_rows(candidate_path)
output_rows: list[dict[str, Any]] = []
missing_candidates = 0
missing_labels = 0
anchor_missing = 0
for row in source_rows:
query_id = row.get(id_field) or row.get("claim_id")
label = row.get("label")
if label is None:
missing_labels += 1
continue
candidate_row = candidates_by_id.get(query_id)
if candidate_row is None:
missing_candidates += 1
continue
evidence = select_evidence(dataset, candidate_row.get("candidates", []), top_k=top_k)
if dataset == "averitec" and input_format == "qa":
evidence = enrich_averitec_qa(evidence, qa_by_evidence or {})
if dataset == "healthver" and not any(item.get("source_type") == "paired_evidence" for item in evidence):
anchor_missing += 1
evidence_summaries = [candidate_summary(item) for item in evidence]
output_rows.append(
{
"id": query_id,
"dataset": dataset,
"split": split,
"claim": row.get("claim", ""),
"label": label,
"input_text": format_input_text(dataset, row.get("claim", ""), evidence, input_format=input_format),
"evidence": evidence_summaries,
"protocol": protocol,
"top_k": top_k,
"input_format": input_format,
}
)
write_jsonl(output_path, output_rows)
return {
"dataset": dataset,
"split": split,
"top_k": top_k,
"input_format": input_format,
"source_rows": len(source_rows),
"output_rows": len(output_rows),
"missing_labels": missing_labels,
"missing_candidates": missing_candidates,
"healthver_anchor_missing": anchor_missing,
"output": str(output_path),
}
def build_dataset(dataset: str, top_k: int, output_root: Path, input_format: str) -> list[dict[str, Any]]:
spec = DATASET_SPECS[dataset]
rows: list[dict[str, Any]] = []
dataset_dir = output_root / dataset
qa_by_evidence: dict[str, dict[str, Any]] = {}
if dataset == "averitec" and input_format == "qa":
qa_by_evidence = load_averitec_qa_by_evidence(Path(spec["qa_evidence"]))
for split, source in spec["splits"].items():
source_path = Path(source)
candidate_path = Path(spec["candidate_template"].format(split=split))
format_suffix = "_qa" if input_format == "qa" else ""
output_path = dataset_dir / f"{split}_top{top_k}{format_suffix}.jsonl"
rows.append(
build_dataset_split(
dataset=dataset,
split=split,
source_path=source_path,
candidate_path=candidate_path,
output_path=output_path,
top_k=top_k,
protocol=spec["protocol"],
id_field=spec["id_field"],
input_format=input_format,
qa_by_evidence=qa_by_evidence,
)
)
return rows
def main() -> None:
parser = argparse.ArgumentParser()
parser.add_argument("--datasets", nargs="*", choices=sorted(DATASET_SPECS), default=sorted(DATASET_SPECS))
parser.add_argument("--top-k", type=int, action="append", required=True)
parser.add_argument("--format", choices=["flat", "qa"], default="flat")
parser.add_argument("--output-root", type=Path, default=Path("outputs/verifier_inputs"))
parser.add_argument("--summary-output", type=Path, default=Path("outputs/stats/verifier_input_summary.csv"))
args = parser.parse_args()
summary_rows: list[dict[str, Any]] = []
for top_k in args.top_k:
for dataset in args.datasets:
summary_rows.extend(
build_dataset(dataset, top_k=top_k, output_root=args.output_root, input_format=args.format)
)
write_csv(args.summary_output, summary_rows)
print(f"Wrote {len(summary_rows)} verifier input split summaries to {args.summary_output}")
if __name__ == "__main__":
main()
|