File size: 7,708 Bytes
d13981b | 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 | #!/usr/bin/env python3
"""Offline G0 audit for deterministic ChartQA full-CoT quality signals."""
from __future__ import annotations
import argparse
import csv
import json
import random
import sys
from collections import Counter
from dataclasses import asdict
from pathlib import Path
from typing import Any, Sequence
ROOT = Path(__file__).resolve().parents[1]
if str(ROOT) not in sys.path:
sys.path.insert(0, str(ROOT))
from reward_utils.chart_cot_verifier import ( # noqa: E402
normalize_reasoning_template,
parse_deplot_table,
summarize_template_diversity,
verify_chart_cot_trajectory,
)
def parse_args(argv: Sequence[str] | None = None) -> argparse.Namespace:
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument("--dataset", required=True)
parser.add_argument("--teacher-jsonl", default="")
parser.add_argument("--out-dir", required=True)
parser.add_argument("--max-samples", type=int, default=0)
parser.add_argument("--seed", type=int, default=42)
return parser.parse_args(argv)
def _load_json(path: Path) -> list[dict[str, Any]]:
data = json.loads(path.read_text(encoding="utf-8"))
if not isinstance(data, list):
raise ValueError(f"Dataset must contain a JSON list: {path}")
return [row for row in data if isinstance(row, dict)]
def _load_jsonl(path: Path) -> list[dict[str, Any]]:
rows: list[dict[str, Any]] = []
for line in path.read_text(encoding="utf-8").splitlines():
if not line.strip():
continue
row = json.loads(line)
if isinstance(row, dict):
rows.append(row)
return rows
def _sample_rows(rows: list[dict[str, Any]], max_samples: int, seed: int) -> list[dict[str, Any]]:
if max_samples <= 0 or max_samples >= len(rows):
return list(rows)
indices = sorted(random.Random(seed).sample(range(len(rows)), max_samples))
return [rows[index] for index in indices]
def _dataset_response(row: dict[str, Any]) -> tuple[str, bool]:
hint = str(row.get("hint") or row.get("visual_fact_hint") or "").strip()
answer = str(row.get("answer") or "").strip()
return f"{hint}\nAnswer: {answer}".strip(), True
def _candidate_response(row: dict[str, Any]) -> tuple[str, bool]:
response = str(row.get("teacher_output") or row.get("response") or "")
answer_correct = bool(row.get("teacher_correct", row.get("answer_correct", False)))
return response, answer_correct
def _join_candidates_to_dataset(
candidates: list[dict[str, Any]],
dataset_rows: list[dict[str, Any]],
) -> list[dict[str, Any]]:
by_image = {
str(row.get("image")): row
for row in dataset_rows
if str(row.get("image") or "").strip()
}
by_question = {
str(row.get("question") or row.get("question_wo_prompt") or "").strip(): row
for row in dataset_rows
if str(row.get("question") or row.get("question_wo_prompt") or "").strip()
}
joined: list[dict[str, Any]] = []
for candidate in candidates:
image = str(candidate.get("image") or "").strip()
question = str(candidate.get("question") or candidate.get("question_wo_prompt") or "").strip()
dataset_row = by_image.get(image) if image else None
if dataset_row is None and question:
dataset_row = by_question.get(question)
joined.append({**(dataset_row or {}), **candidate})
return joined
def _row_result(row: dict[str, Any], response: str, answer_correct: bool, synthesized: bool) -> dict[str, Any]:
deplot = row.get("visual_fact_deplot")
verification = verify_chart_cot_trajectory(response, deplot, answer_correct=answer_correct)
table = parse_deplot_table(deplot)
template = normalize_reasoning_template(verification.parsed.reasoning, table)
return {
"question": str(row.get("question") or row.get("question_wo_prompt") or ""),
"reference_answer": str(row.get("answer") or row.get("reference") or ""),
"response": response,
"synthesized_answer_line": synthesized,
"quality": verification.quality,
"reason_codes": list(verification.reason_codes),
"structure_valid": verification.parsed.structure_valid,
"deplot_available": verification.deplot_available,
"grounded_claims": [asdict(claim) for claim in verification.grounded_claims],
"reasoning_checks": [asdict(check) for check in verification.reasoning_checks],
"conclusion_answer": asdict(verification.conclusion_answer),
"reasoning_template": template,
}
def _summary(rows: list[dict[str, Any]]) -> dict[str, Any]:
qualities = Counter(row["quality"] for row in rows)
quality_counts = {quality: qualities.get(quality, 0) for quality in ("Q0", "Q1", "Q2", "Q3")}
q3_templates = [row["reasoning_template"] for row in rows if row["quality"] == "Q3"]
other_templates = [row["reasoning_template"] for row in rows if row["quality"] != "Q3"]
return {
"sample_count": len(rows),
"quality_counts": quality_counts,
"quality_rates": {
quality: count / max(len(rows), 1) for quality, count in quality_counts.items()
},
"structure_valid_rate": sum(bool(row["structure_valid"]) for row in rows) / max(len(rows), 1),
"deplot_available_rate": sum(bool(row["deplot_available"]) for row in rows) / max(len(rows), 1),
"templates": {
"q3": summarize_template_diversity(q3_templates),
"non_q3": summarize_template_diversity(other_templates),
},
}
def _write_outputs(out_dir: Path, rows: list[dict[str, Any]]) -> None:
out_dir.mkdir(parents=True, exist_ok=True)
summary = _summary(rows)
(out_dir / "chart_cot_quality_summary.json").write_text(
json.dumps(summary, ensure_ascii=False, indent=2, sort_keys=True) + "\n",
encoding="utf-8",
)
with (out_dir / "chart_cot_quality_rows.jsonl").open("w", encoding="utf-8") as handle:
for row in rows:
handle.write(json.dumps(row, ensure_ascii=False, sort_keys=True) + "\n")
with (out_dir / "chart_cot_quality_conflicts.csv").open(
"w", encoding="utf-8", newline=""
) as handle:
writer = csv.DictWriter(
handle,
fieldnames=("quality", "reason_codes", "question", "reference_answer", "response"),
)
writer.writeheader()
for row in rows:
if row["quality"] != "Q0":
continue
writer.writerow(
{
"quality": row["quality"],
"reason_codes": ",".join(row["reason_codes"]),
"question": row["question"],
"reference_answer": row["reference_answer"],
"response": row["response"],
}
)
def main(argv: Sequence[str] | None = None) -> int:
args = parse_args(argv)
dataset_rows = _load_json(Path(args.dataset))
if args.teacher_jsonl:
source_rows = _join_candidates_to_dataset(
_load_jsonl(Path(args.teacher_jsonl)),
dataset_rows,
)
response_builder = _candidate_response
synthesized = False
else:
source_rows = dataset_rows
response_builder = _dataset_response
synthesized = True
sampled = _sample_rows(source_rows, args.max_samples, args.seed)
results = []
for row in sampled:
response, answer_correct = response_builder(row)
results.append(_row_result(row, response, answer_correct, synthesized))
_write_outputs(Path(args.out_dir), results)
return 0
if __name__ == "__main__":
raise SystemExit(main())
|