Datasets:
File size: 6,942 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 | #!/usr/bin/env python3
"""Render the cross-domain manuscript assets from audited saved-response metrics."""
import argparse
from decimal import Decimal, ROUND_HALF_UP
import json
from pathlib import Path
import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt
import numpy as np
from analyze_crossdomain_paper import MODELS, STATES
NAMES = {"qwen38-flash-next": "Qwen 3.8 Flash Next", "gemma4-31b-it": "Gemma 4 31B",
"qwen": "Qwen 3.8 27B", "gemma4-26b-a4b-it": "Gemma 4 26B A4B",
"glm": "GLM 5.3 Flash", "molmo2-8b": "Molmo2-8B"}
def percentage(n, d, places=2):
return str((Decimal(int(n)) * 100 / Decimal(int(d))).quantize(Decimal(10) ** -places, rounding=ROUND_HALF_UP))
def pct(value, places=2):
# Metrics are sample means; remove insignificant summation noise before formatting.
return str(Decimal(str(round(value * 100, 9))).quantize(Decimal(10) ** -places, rounding=ROUND_HALF_UP))
def table(path, header, rows, columns):
path.write_text("\n".join([r"\begin{tabular}{@{}" + columns + r"@{}}", r"\toprule",
header + r" \\", r"\midrule", *rows,
r"\bottomrule", r"\end{tabular}"]) + "\n")
def model_table(data, path):
"""Report complete task success before its decision diagnostics."""
rows = []
for model in MODELS:
cells = []
for source in STATES:
domain = data["domains"][source]
counts = domain["models"][model]["counts"]
cells.extend([
percentage(counts["joint_groups_correct"], 1000, 1),
percentage(counts["groups_correct"], 1000, 1),
percentage(counts["failures"], domain["views"]),
])
rows.append(NAMES[model] + " & " + " & ".join(cells) + r" \\")
rows.extend([
r"\midrule",
r"Always answerable & 0.0 & 0.0 & 40.00 & 0.0 & 0.0 & 25.00 & 0.0 & 0.0 & 33.33 \\",
r"Always unanswerable & 0.0 & 0.0 & 60.00 & 0.0 & 0.0 & 75.00 & 0.0 & 0.0 & 66.67 \\",
])
header = (
r"& \multicolumn{3}{c}{PlotQA} & \multicolumn{3}{c}{CLEVR} & \multicolumn{3}{c}{GQA} \\ "
r"\cmidrule(lr){2-4}\cmidrule(lr){5-7}\cmidrule(l){8-10} "
r"System & $J_5\uparrow$ & $B_5\uparrow$ & $E\downarrow$ "
r"& $J_4\uparrow$ & $B_4\uparrow$ & $E\downarrow$ "
r"& $J_3\uparrow$ & $B_3\uparrow$ & $E\downarrow$"
)
table(path, header, rows, "lrrrrrrrrr")
def main():
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument("--analysis", type=Path, required=True)
parser.add_argument("--output", type=Path, required=True)
args = parser.parse_args()
data = json.loads(args.analysis.read_text())
out = args.output
out.mkdir(parents=True, exist_ok=True)
domains = data["domains"]
model_table(data, out / "model_table.tex")
rows = []
for model in MODELS:
cells = []
for source in STATES:
m = domains[source]["models"][model]["metrics"]
cells.extend([pct(m["balanced_failure"]), pct(m["common_state_failure"])])
rows.append(NAMES[model] + " & " + " & ".join(cells) + r" \\")
rows += [r"\midrule", r"Always answerable & 50.00 & 33.33 & 50.00 & 33.33 & 50.00 & 33.33 \\"]
table(out / "balanced_table.tex", r"System & PlotQA Bal. & $E_{FSM}$ & CLEVR Bal. & $E_{FSM}$ & GQA Bal. & $E_{FSM}$", rows, "lrrrrrr")
rows = []
for model in MODELS:
cells = []
for source in STATES:
c = domains[source]["models"][model]["counts"]
cells.extend([percentage(c["strict_answers_correct"], c["answerable_views"]),
percentage(c["joint_groups_correct"], 1000, 1)])
rows.append(NAMES[model] + " & " + " & ".join(cells) + r" \\")
table(out / "answer_table.tex", r"System & PlotQA $A$ & $J_5$ & CLEVR $A$ & $J_4$ & GQA $A$ & $J_3$", rows, "lrrrrrr")
rows = []
for model in MODELS:
cells = []
for stratum in ["location", "other"]:
c = domains["gqa"]["location_sensitivity"]["models"][model][stratum]
cells.extend([percentage(c["failures"], c["views"]), percentage(c["groups_correct"], c["groups"], 1),
percentage(c["missing_failures"], c["groups"])])
rows.append(NAMES[model] + " & " + " & ".join(cells) + r" \\")
table(out / "location_table.tex", r"System & Loc. Fail & $B_3$ & $M$ Fail & Other Fail & $B_3$ & $M$ Fail", rows, "lrrrrrr")
rows = []
for source in STATES:
for model in MODELS:
d = domains[source]["models"][model]
c = d["counts"]
lo, hi = d["intervals_95"]["failure"]
rows.append(source.upper() + " / " + NAMES[model] + " & " + str(c["valid_decision_errors"]) + "/" + str(c["valid"]) +
" & " + str(c["invalid"]) + " & " + pct(lo) + "--" + pct(hi) + r" \\")
table(out / "denominator_table.tex", r"Domain / system & $W/V$ & Invalid & Failure 95\% interval", rows, "lrrr")
plt.rcParams.update({"font.family": "DejaVu Sans", "font.size": 12, "pdf.fonttype": 42})
fig, axes = plt.subplots(1, 3, figsize=(8.3, 3.7), gridspec_kw={"width_ratios": [5, 4, 3]})
for ax, (source, states) in zip(axes, STATES.items()):
values = np.array([[domains[source]["models"][m]["states"][s]["failures"] / 10 for s in states] for m in MODELS])
ax.imshow(values, cmap="YlOrRd", vmin=0, vmax=80, aspect="auto")
for i in range(len(MODELS)):
for j in range(len(states)):
ax.text(j, i, f"{values[i, j]:.1f}", ha="center", va="center",
color="white" if values[i, j] > 48 else "#222222", fontsize=11)
ax.set_title({"plotqa": "A PlotQA", "clevr": "B CLEVR", "gqa": "C GQA"}[source], loc="left", fontsize=12)
ax.set_xticks(range(len(states)), [{"FULL": "F", "A_SAME": "S", "A_CHANGED": "C", "U_MISSING": "M", "U_INVALID": "I"}[s] for s in states])
ax.set_yticks(range(6), [NAMES[m] for m in MODELS] if source == "plotqa" else [])
ax.tick_params(length=0)
fig.tight_layout(w_pad=.6, rect=(0, .06, 1, 1))
fig.text(.60, .025, "Failure (%); 1,000 views per cell", ha="center", fontsize=12)
for ext in ["pdf", "png"]:
fig.savefig(out / ("state_failures." + ext), dpi=220, bbox_inches="tight")
plt.close(fig)
summary = {"status": "passed", "model_domain_cells": 18,
"rounded_percentages": "decimal half-up from integer counts; percentages are not truncated",
"tables": ["model_table.tex", "balanced_table.tex", "answer_table.tex", "denominator_table.tex", "location_table.tex"],
"figure": "state_failures.pdf", "sha_validation_performed": False}
(out / "asset-validation.json").write_text(json.dumps(summary, indent=2) + "\n")
print(json.dumps(summary))
if __name__ == "__main__":
main()
|