Datasets:
File size: 16,908 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 | #!/usr/bin/env python3
"""Replay the 72,000 saved decisions and build descriptive, paired diagnostics.
Inputs are explicit field projections of final responses, not new model calls.
Existing label and group identifiers are treated as opaque keys. No hashes run.
"""
from __future__ import annotations
import argparse
from collections import Counter, defaultdict
import gzip
import json
from pathlib import Path
import numpy as np
from analyze_crossdomain_visual import strict_answer_match
from analyze_visual_judges import number, valid_parsed
MODELS = ["qwen38-flash-next", "gemma4-31b-it", "qwen", "gemma4-26b-a4b-it", "glm", "molmo2-8b"]
CROSS_KEYS = {m: {"qwen": "qwen3.8-27b-fp8", "glm": "GLM-5.3-Flash"}.get(m, m) for m in MODELS}
STATES = {"plotqa": ["FULL", "A_SAME", "A_CHANGED", "U_MISSING", "U_INVALID"],
"clevr": ["FULL", "A_SAME", "A_CHANGED", "U_MISSING"],
"gqa": ["FULL", "A_SAME", "U_MISSING"]}
COMMON = ["FULL", "A_SAME", "U_MISSING"]
METRICS = ["failure", "group_success", "joint_success", "answer_accuracy_strict",
"balanced_failure", "common_state_failure", "common_state_group_success",
"missing_failure", "answerable_failure", "unanswerable_failure"]
def read(path):
return json.loads(Path(path).read_text())
def rows(path):
with gzip.open(path, "rt") as handle:
return [json.loads(line) for line in handle if line.strip()]
def interval(samples):
return np.quantile(samples, [.025, .975], axis=0).tolist()
def analyze(inputs: Path, expected: Path, draws: int = 10000):
if draws < 10000:
raise ValueError("The manuscript analysis requires at least 10,000 bootstrap draws")
plot_labels = read(inputs / "plotqa/labels.json")
cross_labels = read(inputs / "crossdomain/labels.json")
manifest = [json.loads(line) for line in (inputs / "crossdomain/manifest.jsonl").read_text().splitlines()]
blind = {r["item_id"]: r for r in manifest}
if len(blind) != len(manifest) or len(blind) != 7000:
raise ValueError("Cross-domain manifest must contain 7,000 unique IDs")
plot_manifest = [json.loads(line) for line in (inputs / "plotqa/manifest.jsonl").read_text().splitlines()]
plot_blind = {r["item_id"]: r for r in plot_manifest}
if len(plot_blind) != len(plot_manifest) or len(plot_blind) != 5000:
raise ValueError("PlotQA manifest must contain 5,000 unique IDs")
group_meta = read(inputs / "crossdomain/group-metadata.json")
meta = {r["group_id"]: r for r in group_meta}
if len(meta) != len(group_meta) or len(meta) != 2000:
raise ValueError("Expected 2,000 distinct frozen group metadata records")
label_sets = {"plotqa": plot_labels,
**{s: [r for r in cross_labels if r["source"] == s] for s in ["clevr", "gqa"]}}
old_plot = read(expected / "plotqa-analysis.json")["models"]
old_joint = read(expected / "plotqa-joint.json")["models"]
old_cross = read(expected / "crossdomain-analysis.json")["models"]
outputs = {}
output_rows = 0
for domain in ["plotqa", "crossdomain"]:
expected_ids = {r["item_id"] for r in (plot_labels if domain == "plotqa" else cross_labels)}
questions = plot_blind if domain == "plotqa" else blind
if expected_ids != set(questions):
raise ValueError("Label IDs and frozen question IDs differ")
outputs[domain] = {}
for model in MODELS:
key = model if domain == "plotqa" else CROSS_KEYS[model]
loaded = rows(inputs / domain / (key + ".jsonl.gz"))
records = {r["item_id"]: r for r in loaded}
if len(records) != len(loaded) or set(records) != expected_ids:
raise ValueError(f"{domain}/{model}: duplicate, missing or unexpected IDs")
for r in loaded:
if r["status"] != "completed" and not (
r["status"] == "invalid_schema" and r.get("terminal_invalid") is True):
raise ValueError(f"{domain}/{model}: unfinished response")
if r["question"] != questions[r["item_id"]]["question"]:
raise ValueError("Question differs from frozen blind input")
output_rows += len(loaded)
outputs[domain][model] = records
result = {"schema_version": 1, "models": MODELS, "domains": {}, "provenance": {},
"protocol": {"draws": draws, "seed": 20260915, "unit": "source-question group",
"interval": "descriptive percentile bootstrap; paired within domain; no multiplicity adjustment",
"common_states": COMMON, "inference_calls": 0,
"cross_domain_interpretation": "different questions, states and evidence guarantees; no causal domain effect",
"balanced_failure": "one half answerable failure plus one half unanswerable failure; invalids fail in their truth class",
"new_metrics": "post-hoc descriptive additions; not preregistered",
"source_image_clustering": "one question per source image is the release builder contract; portable labels cannot re-audit original image identities for CLEVR/GQA",
"sha_validation_performed": False}}
total_invalid = 0
for source, labels in label_sets.items():
states = STATES[source]
gold = {r["item_id"]: r for r in labels}
if len(gold) != len(labels) or len(labels) != 1000 * len(states):
raise ValueError(f"{source}: invalid number of unique labels")
grouped = defaultdict(dict)
for r in labels:
if r["state"] not in states or type(r["answerable"]) is not bool:
raise ValueError("Malformed state or target")
if r["answerable"] != (not r["state"].startswith("U_")):
raise ValueError("Inconsistent answerability label")
if r["state"] in grouped[r["group_id"]]:
raise ValueError("Repeated group state")
if source != "plotqa":
g = meta[r["group_id"]]
if g["source"] != source or g["question"] != blind[r["item_id"]]["question"]:
raise ValueError("Release group/question join failed")
grouped[r["group_id"]][r["state"]] = r
if len(grouped) != 1000 or any(set(g) != set(states) for g in grouped.values()):
raise ValueError(f"{source}: incomplete group coverage")
group_ids = sorted(grouped)
ordered = [[grouped[g][s] for s in states] for g in group_ids]
positive = np.array([not s.startswith("U_") for s in states])
common = [states.index(s) for s in COMMON]
vectors = []
domain = {"groups": 1000, "views": len(labels), "states": states, "models": {}}
if source == "gqa":
location = np.array([grouped[g]["FULL"]["target"] in {"left", "right", "top", "bottom"} for g in group_ids])
domain["location_sensitivity"] = {
"selection": "FULL canonical target in {left, right, top, bottom}; all sibling states retained",
"location_groups": int(location.sum()), "other_groups": int((~location).sum()),
"location_group_ids": [g for g, selected in zip(group_ids, location) if selected],
"interpretation": "post-hoc vulnerability stratum, not adjudicated label errors; retained complement is not certified clean",
"models": {}, "intervals": "point estimates only"}
if source != "plotqa":
mm = [meta[g] for g in group_ids]
domain["program_steps"] = dict(sorted(Counter(g["dependency_depth"] for g in mm).items()))
domain["program_steps_note"] = "Stored dependency_depth is len(program steps), not tree depth"
domain["answer_types"] = dict(Counter(g["answer_type"] for g in mm))
for model in MODELS:
records = outputs["plotqa" if source == "plotqa" else "crossdomain"][model]
invalid = np.zeros((1000, len(states)), dtype=bool)
wrong = invalid.copy()
correct_answer = invalid.copy()
for i, gg in enumerate(ordered):
for j, r in enumerate(gg):
prediction = records[r["item_id"]]
if source == "plotqa" and prediction["question"] != records[gg[0]["item_id"]]["question"]:
raise ValueError("PlotQA sibling questions differ")
parsed = valid_parsed(prediction)
invalid[i, j] = parsed is None
wrong[i, j] = parsed is not None and parsed["answerable"] != r["answerable"]
if parsed and parsed["answerable"] and r["answerable"]:
if source == "plotqa":
value, target = number(parsed["answer"]), number(r["target"])
correct_answer[i, j] = value is not None and target is not None and value == target
else:
correct_answer[i, j] = strict_answer_match(parsed["answer"], r["target"])
fail = invalid | wrong
success = ~fail.any(axis=1)
joint = success & correct_answer[:, positive].all(axis=1)
if joint.sum() > success.sum():
raise ValueError("Joint answer success cannot exceed decision success")
pos = fail[:, positive].mean(axis=1)
neg = fail[:, ~positive].mean(axis=1)
matrix = np.column_stack([fail.mean(axis=1), success, joint,
correct_answer[:, positive].mean(axis=1), (pos + neg) / 2,
fail[:, common].mean(axis=1), ~fail[:, common].any(axis=1),
fail[:, states.index("U_MISSING")], pos, neg])
vectors.append(matrix)
counts = {"failures": int(fail.sum()), "invalid": int(invalid.sum()),
"valid_decision_errors": int(wrong.sum()), "valid": int((~invalid).sum()),
"groups_correct": int(success.sum()), "joint_groups_correct": int(joint.sum()),
"strict_answers_correct": int(correct_answer[:, positive].sum()),
"answerable_views": int(1000 * positive.sum()),
"unanswerable_views": int(1000 * (~positive).sum())}
state_counts = {s: {"failures": int(fail[:, j].sum()), "invalid": int(invalid[:, j].sum()),
"valid_decision_errors": int(wrong[:, j].sum()), "views": 1000}
for j, s in enumerate(states)}
if source == "plotqa":
old = old_plot[model]
checks = [(counts["failures"], old["error_fraction_if_invalid_count_wrong"]["numerator"]),
(counts["invalid"], old["invalid_or_missing"]),
(counts["valid_decision_errors"], old["answerability_errors"]["numerator"]),
(counts["groups_correct"], old["five_view_groups_all_correct"]["numerator"]),
(counts["joint_groups_correct"], old_joint[model]["strict_joint_all_five_groups"]["numerator"])]
for s in states:
checks.append((state_counts[s]["failures"], old["state_counts"][s]["failures_including_invalid"]["numerator"]))
else:
old = old_cross[CROSS_KEYS[model]]["by_source"][source]
checks = [(counts["failures"], old["failures_full_denominator"]["numerator"]),
(counts["invalid"], old["invalid_outputs"]),
(counts["valid_decision_errors"], old["answerability_errors"]["numerator"]),
(counts["groups_correct"], old["groups_all_states_correct"]["numerator"]),
(counts["strict_answers_correct"], old["answer_correct_on_all_answerable_strict_lower_bound"]["numerator"])]
for s in states:
checks.append((state_counts[s]["failures"], old["states"][s]["failures_full_denominator"]["numerator"]))
if any(a != b for a, b in checks):
raise ValueError(f"Saved aggregate mismatch: {source}/{model}: {checks}")
domain["models"][model] = {"counts": counts, "states": state_counts,
"metrics": dict(zip(METRICS, matrix.mean(axis=0).tolist())),
"saved_aggregate_checks": len(checks),
"invalid_item_ids": [r["item_id"] for i, gg in enumerate(ordered)
for j, r in enumerate(gg) if invalid[i, j]]}
if source == "gqa":
domain["location_sensitivity"]["models"][model] = {}
for stratum, mask in [("location", location), ("other", ~location)]:
domain["location_sensitivity"]["models"][model][stratum] = {
"groups": int(mask.sum()), "views": int(mask.sum() * len(states)),
"failures": int(fail[mask].sum()), "groups_correct": int(success[mask].sum()),
"missing_failures": int(fail[mask, states.index("U_MISSING")].sum()),
"invalid": int(invalid[mask].sum())}
total_invalid += counts["invalid"]
if model == "qwen" and source != "plotqa":
backend = {"subsets": {}, "group_composition": dict(Counter(
"mixed" if len({records[r["item_id"]]["saved_backend"] for r in gg}) > 1
else records[gg[0]["item_id"]]["saved_backend"] for gg in ordered))}
for name in ["bridge", "simflow"]:
mask = np.array([[records[r["item_id"]]["saved_backend"] == name for r in gg] for gg in ordered])
backend["subsets"][name] = {"views": int(mask.sum()), "failures": int(fail[mask].sum()),
"invalid": int(invalid[mask].sum()), "state_counts": {s: int(mask[:, j].sum()) for j, s in enumerate(states)}}
backend["interpretation"] = "Availability-selected disjoint subsets; not a matched backend comparison"
result["provenance"][source] = backend
values = np.stack(vectors, axis=1)
rng = np.random.default_rng(20260915)
boot = np.empty((draws, len(MODELS), len(METRICS)))
for start in range(0, draws, 100):
end = min(start + 100, draws)
selected = rng.integers(0, 1000, size=(end - start, 1000))
boot[start:end] = values[selected].mean(axis=1)
for i, model in enumerate(MODELS):
domain["models"][model]["intervals_95"] = {m: interval(boot[:, i, k]) for k, m in enumerate(METRICS)}
contrasts = {}
for i, left in enumerate(MODELS):
for j in range(i + 1, len(MODELS)):
right = MODELS[j]
contrasts[left + "__" + right] = {"left": left, "right": right,
**{metric: {"left_minus_right_pp": float((values[:, i, k] - values[:, j, k]).mean() * 100),
"interval_95_pp": interval((boot[:, i, k] - boot[:, j, k]) * 100)}
for k, metric in enumerate(METRICS)}}
domain["paired_contrasts"] = contrasts
domain["analytic_baselines"] = {"always_answerable_failure": float((~positive).mean()),
"always_unanswerable_failure": float(positive.mean()), "constant_decision_group_success": 0,
"constant_decision_balanced_failure": .5}
result["domains"][source] = domain
result["validation"] = {"status": "passed", "selected_model_responses": output_rows,
"unique_input_views": 12000, "source_question_groups": 3000, "model_domain_cells": 18,
"terminal_invalid_responses": total_invalid, "valid_responses": output_rows - total_invalid,
"missing_or_duplicate_rows": 0, "original_aggregate_counts_reproduced": True,
"all_model_frozen_question_joins": True,
"output_projection_scope": "parsed response fields, not independent raw-output/token validation",
"sha_validation_performed": False}
return result
def main():
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument("--inputs", type=Path, required=True)
parser.add_argument("--expected", type=Path, required=True)
parser.add_argument("--output", type=Path, required=True)
args = parser.parse_args()
result = analyze(args.inputs, args.expected)
args.output.parent.mkdir(parents=True, exist_ok=True)
args.output.write_text(json.dumps(result, ensure_ascii=False, indent=2, allow_nan=False) + "\n")
print(json.dumps(result["validation"]))
if __name__ == "__main__":
main()
|