Datasets:
File size: 3,977 Bytes
e1ced61 | 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 | #!/usr/bin/env python3
"""Reproduce all 18 cells and paired intervals from 72,000 saved scoring records."""
from __future__ import annotations
import argparse
import json
import subprocess
import sys
from pathlib import Path
from analyze_crossdomain_paper import CROSS_KEYS, MODELS, analyze
from hf_release_common import read_json, read_rows, write_json
from score_hf_dataset import score
def replay(dataset, output):
output.mkdir(parents=True, exist_ok=True)
# Histogram keys are integers in memory and strings in the saved JSON.
# Compare the same serialized representation used by the original replay.
result = json.loads(
json.dumps(analyze(dataset / "inputs", dataset / "expected"), allow_nan=False)
)
expected = read_json(dataset / "expected/revision-analysis.json")
if result != expected:
differing = [key for key in result if result[key] != expected.get(key)]
raise ValueError("Numerical replay differs from frozen analysis: " + ", ".join(differing))
labels = read_rows(dataset / "metadata/views.jsonl.gz")
for model in MODELS:
predictions = read_rows(dataset / f"inputs/plotqa/{model}.jsonl.gz") + read_rows(
dataset / f"inputs/crossdomain/{CROSS_KEYS[model]}.jsonl.gz"
)
scored = score(labels, predictions)
for source, values in scored["by_source"].items():
archived = result["domains"][source]["models"][model]
if values["counts"] != archived["counts"] or values["states"] != archived["states"]:
raise ValueError(
"Public prediction scorer disagrees with paper: " + source + "/" + model
)
write_json(output / (model + "-score.json"), scored)
from build_answerability_rewrite_assets import build as build_chart
result_path = output / "analysis.json"
write_json(result_path, result)
subprocess.run(
[
sys.executable,
str(dataset / "scripts/build_crossdomain_paper_assets.py"),
"--analysis",
str(result_path),
"--output",
str(output / "tables"),
],
check=True,
)
build_chart(
dataset / "inputs",
result_path,
output / "charts",
dataset / "evidence/plotqa/bounded-witness-proofs.jsonl.gz",
)
from collections import Counter
finite = read_rows(dataset / "evidence/finite/original-results.jsonl.gz")
solver = read_rows(dataset / "evidence/finite/solver-results.jsonl.gz")
if (
len(finite) != 1024
or len(solver) != 1456
or Counter(row["truth"] for row in solver) != {"ambiguous": 822, "constant": 634}
):
raise ValueError("Finite study counts differ")
if not all(row["enumeration_status"] == row["smt_status"] == row["truth"] for row in solver):
raise ValueError("Archived solver decisions disagree")
report = {
"status": "passed",
"saved_records": 72000,
"model_domain_cells": 18,
"exact_analysis_replay": True,
"bootstrap_draws": 10000,
"new_prediction_scorer_matches_all_18_cells": True,
"finite_cases": len(finite),
"solver_cases": len(solver),
"solver_constructors_rerun": False,
"inference_calls": 0,
"gqa_question_text": "ID references; original text verified separately during reconstruction",
}
write_json(output / "replay-validation.json", report)
return report
if __name__ == "__main__":
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument("--dataset", type=Path, default=Path(__file__).resolve().parents[1])
parser.add_argument("--output", type=Path, required=True)
args = parser.parse_args()
if args.output.resolve().is_relative_to(args.dataset.resolve()):
parser.error("Write replay outputs outside the frozen dataset directory")
print(json.dumps(replay(args.dataset, args.output)))
|