File size: 73,248 Bytes
3ccaf5a | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694 695 696 697 698 699 700 701 702 703 704 705 706 707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864 865 866 867 868 869 870 871 872 873 874 875 876 877 878 879 880 881 882 883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076 1077 1078 1079 1080 1081 1082 1083 1084 1085 1086 1087 1088 1089 1090 1091 1092 1093 1094 1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437 1438 1439 1440 1441 1442 1443 1444 1445 1446 1447 1448 1449 1450 1451 1452 1453 1454 1455 1456 1457 1458 1459 1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491 1492 1493 1494 1495 1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541 1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562 1563 1564 1565 1566 1567 1568 1569 1570 1571 1572 1573 1574 1575 1576 1577 1578 1579 1580 1581 1582 1583 1584 1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599 1600 1601 1602 1603 1604 1605 1606 1607 1608 1609 1610 1611 1612 1613 1614 1615 1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 | """Generate the headline figures and numerics table for the paper.
Design conventions follow what top-venue ML papers (NeurIPS, ICLR, Nature MI)
actually ship, not what matplotlib defaults give you. Specifically:
- Base font 8pt, axis labels 9pt, tick labels 7pt, legend 7pt.
- Tableau-10 muted palette for categorical series; Nature-muted for
error-type stacks.
- CI band alpha 0.08 with a 0.4-alpha 0.5pt edge, NOT the chunky 0.12
fills that make overlapping series look like mud.
- Solid fills + 0.8pt white edges on stacked bars. NO hatching.
- No end-caps on forest-plot whiskers; 4pt filled-circle markers.
- Zero reference line solid grey (#999999), not dashed.
- Type-42 fonts so the typesetter can re-kern; Type-3 is an amateur tell.
- NeurIPS widths: 3.25" single-column, 6.75" double-column.
Data consumed:
results/metrics/{model}_{quant}_{bench}_run0_metrics.json point estimates
results/metrics/{model}_{quant}_{bench}_run0_ci.json bootstrap CIs
results/metrics/{model}_{quant}_{bench}_run0_sig.json paired sig tests
"""
import argparse
import glob
import json
import os
import re
import sys
from collections import defaultdict
from typing import Dict, List, Optional, Tuple
import matplotlib.pyplot as plt
import matplotlib as mpl
import numpy as np
sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
# ---------------------------------------------------------------------------
# Global style (Q1-venue defaults)
# ---------------------------------------------------------------------------
mpl.rcParams.update({
"font.family": "sans-serif",
"font.sans-serif": ["Inter", "Helvetica Neue", "Arial", "DejaVu Sans"],
"font.size": 8,
"axes.titlesize": 9,
"axes.titleweight": "regular",
"axes.titlepad": 6,
"axes.labelsize": 9,
"axes.labelpad": 3,
"xtick.labelsize": 7,
"ytick.labelsize": 7,
"legend.fontsize": 7,
"legend.frameon": False,
"figure.dpi": 200,
"savefig.dpi": 400,
"savefig.bbox": "tight",
"pdf.fonttype": 42,
"ps.fonttype": 42,
"axes.linewidth": 0.6,
"axes.edgecolor": "#333333",
"axes.labelcolor": "#222222",
"axes.titlecolor": "#222222",
"xtick.color": "#333333",
"ytick.color": "#333333",
"xtick.major.width": 0.6,
"ytick.major.width": 0.6,
"xtick.major.size": 3,
"ytick.major.size": 3,
"xtick.major.pad": 2,
"ytick.major.pad": 2,
"axes.spines.top": False,
"axes.spines.right": False,
"grid.color": "#EAEAEA",
"grid.linewidth": 0.5,
"grid.linestyle": "-",
"lines.linewidth": 1.3,
"lines.solid_capstyle": "round",
"patch.linewidth": 0.0,
"hatch.linewidth": 0.0, # we don't hatch
})
# Tableau-10 muted: a de-facto standard for categorical ML figures.
# Darker shade = base quantization, paired lighter shade = restored variant.
METHOD_COLOR = {
"awq_w4": "#4E79A7",
"awq_w4_restored": "#A0CBE8",
"gptq_w4": "#E15759",
"gptq_w4_restored": "#FF9D9A",
"bnb_nf4_w4": "#59A14F",
"bnb_nf4_w4_restored": "#8CD17D",
}
METHOD_ORDER = ["awq_w4", "gptq_w4", "bnb_nf4_w4"]
METHOD_PRETTY = {"awq_w4": "AWQ w4", "gptq_w4": "GPTQ w4", "bnb_nf4_w4": "BnB NF4"}
# Nature-muted error-type palette — categorical-ish but visually calm.
ERROR_COLORS = {
"conceptual": "#264653",
"methodological": "#2A9D8F",
"executional": "#E9C46A",
"logical": "#E76F51",
}
ERROR_TYPES = ["conceptual", "methodological", "executional", "logical"]
BENCHMARKS = ["gsm8k", "math500", "gpqa"]
BENCH_PRETTY = {"gsm8k": "GSM8K", "math500": "MATH-500", "gpqa": "GPQA-Diamond"}
GREY_REF = "#999999" # zero/reference line
GREY_LIGHT = "#C7C7C7"
GREY_TEXT = "#555555"
# ---------------------------------------------------------------------------
# Data loading
# ---------------------------------------------------------------------------
def _benchmark_from_filename(fname: str) -> str:
for bench in BENCHMARKS:
if re.search(rf"_{bench}_run\d+_metrics\.json$", fname):
return bench
return "unknown"
def _parse_ci_name(fname: str, suffix: str):
base = fname.replace(suffix, "")
m = re.match(r"(.+?)_(awq_w\d+(?:_restored)?|gptq_w\d+(?:_restored)?|bnb_nf\d+_w\d+(?:_restored)?)_(\w+)_run\d+$", base)
if not m:
return None
return m.group(1), m.group(2), m.group(3)
def load_all(metrics_dir: str) -> Dict[Tuple[str, str, str], dict]:
data: Dict[Tuple[str, str, str], dict] = {}
for f in glob.glob(os.path.join(metrics_dir, "*_metrics.json")):
with open(f) as fp:
d = json.load(fp)
bench = _benchmark_from_filename(os.path.basename(f))
model = d.get("model")
quant = d.get("quantization")
if model and quant:
data.setdefault((model, quant, bench), {}).update(d)
for f in glob.glob(os.path.join(metrics_dir, "*_ci.json")):
parsed = _parse_ci_name(os.path.basename(f), "_ci.json")
if not parsed:
continue
model, quant, bench = parsed
if bench not in BENCHMARKS:
continue
with open(f) as fp:
d = json.load(fp)
data.setdefault((model, quant, bench), {})["ci"] = d
for f in glob.glob(os.path.join(metrics_dir, "*_sig.json")):
parsed = _parse_ci_name(os.path.basename(f), "_sig.json")
if not parsed:
continue
model, quant, bench = parsed
if bench not in BENCHMARKS:
continue
with open(f) as fp:
d = json.load(fp)
data.setdefault((model, quant, bench), {})["sig_vs_restored"] = d
return data
def sig_mark(p: Optional[float]) -> str:
"""Single-asterisk convention; threshold documented in the legend footnote."""
if p is None:
return ""
return "*" if p < 0.05 else ""
def sig_stars_tex(p: Optional[float]) -> str:
"""For LaTeX table only — three-tier stars since the table has room."""
if p is None:
return ""
if p < 0.001:
return "$^{***}$"
if p < 0.01:
return "$^{**}$"
if p < 0.05:
return "$^{*}$"
return ""
# ---------------------------------------------------------------------------
# Fig 1 — SSR curves with CI bands
# ---------------------------------------------------------------------------
def fig_paper_1_ssr(data, primary_model: str, output_path: str):
"""Step-survival with 95% bootstrap CI bands. One panel per benchmark.
Base = solid + filled CI band (alpha 0.08 with a 0.5pt edge at 0.35
alpha — the edge keeps the band from dissolving into the other bands).
Restored = dashed line only, no band (showing 6 bands would be mud).
"""
fig, axes = plt.subplots(1, 3, figsize=(6.75, 2.15),
sharey=True, constrained_layout=True)
for ax, bench in zip(axes, BENCHMARKS):
max_d = 0
for method in METHOD_ORDER:
color = METHOD_COLOR[method]
color_rest = METHOD_COLOR[method + "_restored"]
base_entry = data.get((primary_model, method, bench)) or {}
rest_entry = data.get((primary_model, method + "_restored", bench)) or {}
# Base: line + CI band + edge
ssr = base_entry.get("ssr_curve", [])
if ssr:
x = np.arange(len(ssr))
max_d = max(max_d, len(ssr))
ci = (base_entry.get("ci") or {}).get("ssr_curve_ci", {}) or {}
lo = ci.get("ci_lo") or []
hi = ci.get("ci_hi") or []
if lo and hi:
lo_arr = np.array([np.nan if v is None else v for v in lo[: len(ssr)]])
hi_arr = np.array([np.nan if v is None else v for v in hi[: len(ssr)]])
valid = ~(np.isnan(lo_arr) | np.isnan(hi_arr))
if valid.any():
ax.fill_between(x[valid], lo_arr[valid], hi_arr[valid],
color=color, alpha=0.08, linewidth=0, zorder=1)
ax.plot(x[valid], lo_arr[valid], color=color,
linewidth=0.5, alpha=0.35, zorder=2)
ax.plot(x[valid], hi_arr[valid], color=color,
linewidth=0.5, alpha=0.35, zorder=2)
ax.plot(x, ssr, "-", color=color, linewidth=1.3,
label=METHOD_PRETTY[method], zorder=4)
# Restored: dashed line only (keep the plot readable)
ssr_r = rest_entry.get("ssr_curve", [])
if ssr_r:
xr = np.arange(len(ssr_r))
max_d = max(max_d, len(ssr_r))
ax.plot(xr, ssr_r, "--", color=color_rest, linewidth=1.1,
label=METHOD_PRETTY[method] + " (rest.)", zorder=3)
ax.set_xlim(0, max(12, min(max_d, 25)))
ax.set_ylim(0, 1.02)
ax.set_xlabel("Reasoning step depth")
ax.set_yticks([0, 0.25, 0.5, 0.75, 1.0])
ax.yaxis.grid(True)
ax.set_axisbelow(True)
ax.text(0.98, 0.96, BENCH_PRETTY[bench], transform=ax.transAxes,
ha="right", va="top", fontsize=7.5, color=GREY_TEXT)
axes[0].set_ylabel("Step survival rate")
handles, labels = axes[0].get_legend_handles_labels()
seen = set()
uniq = [(h, l) for h, l in zip(handles, labels) if not (l in seen or seen.add(l))]
if uniq:
h2, l2 = zip(*uniq)
fig.legend(h2, l2, loc="lower center", ncol=min(6, len(l2)),
bbox_to_anchor=(0.5, -0.09),
columnspacing=1.6, handlelength=2.4, handletextpad=0.6)
fig.savefig(output_path)
plt.close(fig)
print(f" Paper fig 1 saved: {output_path}")
# ---------------------------------------------------------------------------
# Fig 2 — Error counts per 100 problems, stacked (one focus benchmark)
# ---------------------------------------------------------------------------
def fig_paper_2_error_mix(data, primary_model: str, primary_benchmark: str, output_path: str):
"""Single focused panel: errors per 100 problems for the primary
(model, benchmark), stacked by error type, with quantized vs restored
bars side-by-side for each method.
White edges between stack segments (DeepMind/Anthropic style) — no hatching.
Each pair carries a small Δ annotation showing the post-restoration change
in total errors, so the reader can see direction at a glance even when
bar heights look near-identical.
"""
# Slightly wider canvas so the per-bar variant labels ("Quant.", "Rest.")
# and the Δ annotations have room without crowding the method names.
fig, ax = plt.subplots(figsize=(4.2, 2.9), constrained_layout=True)
x = np.arange(len(METHOD_ORDER), dtype=float)
width = 0.32
positions = {
"": x - width / 2 - 0.02,
"_restored": x + width / 2 + 0.02,
}
max_total = 0.0
totals_by = {}
for suffix in ("", "_restored"):
bottoms = np.zeros(len(METHOD_ORDER))
totals = np.zeros(len(METHOD_ORDER))
for e_idx, etype in enumerate(ERROR_TYPES):
heights = []
for method in METHOD_ORDER:
q = method + suffix
entry = data.get((primary_model, q, primary_benchmark)) or {}
acc = entry.get("accuracy") or 0.0
dist = entry.get("error_type_dist") or {}
h = (1 - acc) * 100 * dist.get(etype, 0)
heights.append(h)
heights = np.array(heights)
totals += heights
ax.bar(positions[suffix], heights, width=width, bottom=bottoms,
color=ERROR_COLORS[etype], edgecolor="white", linewidth=0.8,
label=etype.capitalize() if suffix == "" else None)
bottoms += heights
totals_by[suffix] = totals
max_total = max(max_total, totals.max())
# Top-of-bar total labels — show one decimal so a 0.2-pp change isn't
# rounded into invisibility (the previous "28 / 28" was misleading).
for suffix in ("", "_restored"):
for mi, t in enumerate(totals_by[suffix]):
if t > 0:
ax.text(positions[suffix][mi], t + max_total * 0.025,
f"{t:.1f}",
ha="center", va="bottom", fontsize=6.8,
color="#222222")
# Δ annotation centered between each (Quant., Rest.) pair, sitting just
# above the taller of the two bars. Green ↓ = restoration reduced errors;
# red ↑ = restoration made it worse. The arrow encodes direction so the
# reader doesn't have to compare two near-identical heights by eye.
for mi in range(len(METHOD_ORDER)):
q_total = totals_by[""][mi]
r_total = totals_by["_restored"][mi]
delta = r_total - q_total
if abs(delta) < 1e-3:
continue
# Sit the badge above the column-pair, clear of the per-bar totals.
y_badge = max(q_total, r_total) + max_total * 0.115
improved = delta < 0
arrow = "↓" if improved else "↑"
color = "#2A9D74" if improved else "#C9534F"
ax.text(x[mi], y_badge,
f"{arrow} {abs(delta):.1f}",
ha="center", va="bottom", fontsize=6.6,
color=color, fontweight="bold")
# Sub-label under each bar identifying the variant. Spelled out so the
# figure is self-contained — "Q" / "R" alone forced the reader to chase
# an off-figure key.
for mi, method in enumerate(METHOD_ORDER):
ax.text(positions[""][mi], -max_total * 0.04, "Quant.",
ha="center", va="top", fontsize=6.4, color=GREY_TEXT)
ax.text(positions["_restored"][mi], -max_total * 0.04, "Rest.",
ha="center", va="top", fontsize=6.4, color=GREY_TEXT)
ax.set_xticks(x)
ax.set_xticklabels([METHOD_PRETTY[m] for m in METHOD_ORDER])
# Extra pad so method labels sit clearly below the Quant./Rest. row.
ax.tick_params(axis="x", which="major", pad=14)
# Headroom for the Δ badge above the tallest bar.
ax.set_ylim(0, max_total * 1.25 + 1)
ax.set_ylabel("Errors per 100 problems")
ax.yaxis.grid(True)
ax.set_axisbelow(True)
# Right-top panel tag (benchmark + model); keep it small and grey.
ax.text(0.98, 0.97,
f"{primary_model} · {BENCH_PRETTY[primary_benchmark]}",
transform=ax.transAxes, ha="right", va="top",
fontsize=7, color=GREY_TEXT)
# Legend: error types beneath the x-axis labels.
ax.legend(loc="upper center", bbox_to_anchor=(0.5, -0.18),
ncol=4, handlelength=1.2, columnspacing=1.3, handletextpad=0.5,
frameon=False)
fig.savefig(output_path)
plt.close(fig)
print(f" Paper fig 2 saved: {output_path}")
# ---------------------------------------------------------------------------
# Fig 3 — Forest plot of ΔAccuracy
# ---------------------------------------------------------------------------
def fig_paper_3_forest(data, models: List[str], output_path: str):
"""Forest plot: ΔAccuracy with 95% paired-bootstrap CI whiskers.
One panel per benchmark. Row = (model, method). Marker = filled circle
4pt, no end-caps on whiskers, zero line solid #999999, single asterisk
at right of rows where p<.05 with a footnote explaining.
"""
# Skip any model that has no (base, restored) paired-sig data in any
# (method, benchmark) cell — otherwise those rows render as empty space
# inside every panel and make the plot look broken.
def _has_any_sig(model):
for method in METHOD_ORDER:
for bench in BENCHMARKS:
sig = (data.get((model, method, bench)) or {}).get("sig_vs_restored") or {}
if sig and sig.get("n_pairs", 0) > 0:
return True
return False
models = [m for m in models if _has_any_sig(m)]
if not models:
print(" [SKIP] fig 3: no paired sig data for any model")
return
rows: List[Tuple[str, str]] = [(m, q) for m in models for q in METHOD_ORDER]
n_rows = len(rows)
row_height = 0.32
fig_h = row_height * n_rows + 1.1
fig, axes = plt.subplots(1, 3, figsize=(6.75, fig_h),
sharex=True, sharey=True, constrained_layout=True)
# Compute shared x range for symmetry about 0.
all_bounds = []
for (model, method) in rows:
for bench in BENCHMARKS:
sig = (data.get((model, method, bench)) or {}).get("sig_vs_restored") or {}
if sig and sig.get("n_pairs", 0) > 0:
lo = sig.get("delta_acc_ci_lo", 0) * 100
hi = sig.get("delta_acc_ci_hi", 0) * 100
all_bounds += [lo, hi]
if all_bounds:
bound = max(abs(min(all_bounds)), abs(max(all_bounds)))
xlim = (-bound * 1.12 - 2, bound * 1.12 + 2)
else:
xlim = (-20, 20)
for ax, bench in zip(axes, BENCHMARKS):
ax.axvline(0, color=GREY_REF, linewidth=0.6, zorder=1)
for row_idx, (model, method) in enumerate(rows):
y = n_rows - 1 - row_idx
entry = data.get((model, method, bench)) or {}
sig = entry.get("sig_vs_restored") or {}
if not sig or sig.get("n_pairs", 0) == 0:
continue
delta = sig.get("delta_acc_observed", 0) * 100
lo = sig.get("delta_acc_ci_lo", delta / 100) * 100
hi = sig.get("delta_acc_ci_hi", delta / 100) * 100
p = sig.get("p_value")
color = METHOD_COLOR[method]
ax.plot([lo, hi], [y, y], color=color, linewidth=1.0, zorder=2, solid_capstyle="butt")
ax.plot(delta, y, "o", markersize=4, color=color,
markeredgewidth=0, zorder=3)
# Place the sig star just to the right of the CI whisker's high
# end — visually attached to the data point rather than parked at
# the panel edge.
star = sig_mark(p)
if star:
ax.text(hi + (xlim[1] - xlim[0]) * 0.02, y, star,
va="center", ha="left",
fontsize=9, color=color, fontweight="bold")
# Faint horizontal separators between model blocks.
for i in range(1, len(models)):
y_sep = n_rows - i * len(METHOD_ORDER) - 0.5
ax.axhline(y_sep, color=GREY_LIGHT, linewidth=0.3, zorder=0)
ax.set_yticks([n_rows - 1 - i for i in range(n_rows)])
ax.set_yticklabels([METHOD_PRETTY[rows[i][1]] for i in range(n_rows)])
ax.set_xlim(*xlim)
ax.xaxis.grid(True)
ax.set_axisbelow(True)
ax.set_xlabel(r"$\Delta$Accuracy (pp)")
# Corner benchmark tag
ax.text(0.98, 1.0, BENCH_PRETTY[bench], transform=ax.transAxes,
ha="right", va="bottom", fontsize=7.5, color=GREY_TEXT)
# Model labels on the left, vertically centered on each model block.
block = len(METHOD_ORDER)
for i, model in enumerate(models):
y_mid = n_rows - 1 - (i * block + (block - 1) / 2)
axes[0].text(-0.46, y_mid, model,
transform=axes[0].get_yaxis_transform(),
ha="right", va="center", fontsize=8, color="#222222")
# Bottom-left significance-key footnote.
fig.text(0.02, -0.02, r"$*\,p<.05$ (paired bootstrap, 5000 iter.)",
ha="left", va="top", fontsize=6.5, color=GREY_TEXT)
fig.savefig(output_path)
plt.close(fig)
print(f" Paper fig 3 saved: {output_path}")
# ---------------------------------------------------------------------------
# Table — LaTeX numerics
# ---------------------------------------------------------------------------
def table_paper_tex(data, models: List[str], output_path: str):
def fmt_pct(x, ci=None):
if x is None:
return "--"
s = f"{x * 100:.1f}"
if ci and ci[0] is not None and ci[1] is not None:
s += f"\\,[{ci[0] * 100:.1f},{ci[1] * 100:.1f}]"
return s
def fmt_f(x, ci=None, d=2):
if x is None or x == float("inf"):
return "--"
s = f"{x:.{d}f}"
if ci and ci[0] is not None and ci[1] is not None:
s += f"\\,[{ci[0]:.{d}f},{ci[1]:.{d}f}]"
return s
def ci_pair(entry, field):
ci = (entry.get("ci") or {}).get(field)
if not ci:
return None
return ci.get("ci_lo"), ci.get("ci_hi")
rows = []
for model in models:
for bench in BENCHMARKS:
for method in METHOD_ORDER:
base = data.get((model, method, bench)) or {}
rest = data.get((model, method + "_restored", bench)) or {}
if not base and not rest:
continue
sig = base.get("sig_vs_restored") or {}
p = sig.get("p_value")
d_acc_obs = sig.get("delta_acc_observed")
rows.append({
"model": model,
"bench": BENCH_PRETTY[bench],
"method": METHOD_PRETTY[method],
"acc_base": fmt_pct(base.get("accuracy"), ci_pair(base, "accuracy")),
"acc_rest": fmt_pct(rest.get("accuracy"), ci_pair(rest, "accuracy")),
"d_acc": ("--" if d_acc_obs is None
else f"{d_acc_obs * 100:+.1f}{sig_stars_tex(p)}"),
"ffs_base": fmt_f(base.get("avg_ffs"), ci_pair(base, "avg_ffs")),
"ffs_rest": fmt_f(rest.get("avg_ffs"), ci_pair(rest, "avg_ffs")),
"ecr_base": fmt_pct(base.get("ecr"), ci_pair(base, "ecr")),
"ecr_rest": fmt_pct(rest.get("ecr"), ci_pair(rest, "ecr")),
})
# Wrap the tabular in \resizebox so the table never overflows the text
# width regardless of how many digits the CIs produce.
header = (
"\\begin{table}[t]\n"
"\\centering\n"
"\\caption{Per-(model, benchmark, method) results. Point estimates followed by 95\\% bootstrap CIs "
"in brackets. Accuracy and ECR in \\%, FFS in step index. $\\Delta$Acc is restored~--~base accuracy "
"(pp); paired-bootstrap significance: $^{*}p<.05$, $^{**}p<.01$, $^{***}p<.001$.}\n"
"\\label{tab:main}\n"
"\\resizebox{\\textwidth}{!}{%\n"
"\\begin{tabular}{@{}lllcccccc@{}}\n"
"\\toprule\n"
"Model & Benchmark & Method "
"& Acc$_\\mathrm{base}$ & Acc$_\\mathrm{rest}$ & $\\Delta$Acc "
"& FFS$_\\mathrm{base}\\!\\to\\!\\mathrm{rest}$ "
"& ECR$_\\mathrm{base}\\!\\to\\!\\mathrm{rest}$ \\\\\n"
"\\midrule\n"
)
body_lines = []
last_model = None
last_bench = None
for row in rows:
model_cell = row["model"] if row["model"] != last_model else ""
bench_cell = row["bench"] if (row["model"] != last_model or row["bench"] != last_bench) else ""
if model_cell and last_model is not None:
body_lines.append("\\addlinespace[2pt]")
body_lines.append(
f"{model_cell} & {bench_cell} & {row['method']} "
f"& {row['acc_base']} & {row['acc_rest']} & {row['d_acc']} "
f"& {row['ffs_base']}\\,$\\to$\\,{row['ffs_rest']} "
f"& {row['ecr_base']}\\,$\\to$\\,{row['ecr_rest']} \\\\"
)
last_model = row["model"]
last_bench = row["bench"]
footer = "\\bottomrule\n\\end{tabular}%\n}\n\\end{table}\n"
with open(output_path, "w") as f:
f.write(header + "\n".join(body_lines) + "\n" + footer)
print(f" Paper table saved: {output_path}")
# ---------------------------------------------------------------------------
# Per-problem diagnosis loader (used by fig 4 and fig 5)
# ---------------------------------------------------------------------------
def _load_diagnosis(diagnosis_dir: str, model: str, quant: str, benchmark: str) -> List[dict]:
path = os.path.join(diagnosis_dir, quant, model, f"{benchmark}_run0.jsonl")
if not os.path.exists(path):
return []
out = []
with open(path) as f:
for line in f:
out.append(json.loads(line))
return out
def _first_failure_step(steps: List[dict]) -> Optional[int]:
for s in steps:
if s.get("is_correct") is False:
return s.get("index")
return None
def _load_segmented_fp16(segmented_dir: str, model: str, benchmark: str) -> List[dict]:
path = os.path.join(segmented_dir, "fp16", model, f"{benchmark}_run0.jsonl")
if not os.path.exists(path):
return []
out = []
with open(path) as f:
for line in f:
out.append(json.loads(line))
return out
# ---------------------------------------------------------------------------
# Fig 0 — Method schematic / pipeline diagram
# ---------------------------------------------------------------------------
def fig_paper_0_pipeline(output_path: str):
"""Method schematic — four coloured stage bands with mini visual
metaphors (step-dot ribbons, a DTW alignment glyph, an error-type donut,
metric badges) so the figure summarises the method without needing to
read the body text.
Stages:
1. Generation — FP16 + quantized CoTs
2. Diagnosis — DTW alignment, per-step scoring, step-level metrics
3. Silver-bullet — failures sampled across the 4 error types
4. Restoration — QLoRA on the quantized base → restored model
"""
from matplotlib.patches import (FancyBboxPatch, FancyArrowPatch,
Rectangle, Circle, Wedge)
# Wider, slightly taller canvas so the new mini-visuals breathe.
fig, ax = plt.subplots(figsize=(8.6, 4.5))
ax.set_xlim(0, 100)
ax.set_ylim(2, 74)
ax.set_axis_off()
# -------- Palette ------------------------------------------------------
C_FP16 = METHOD_COLOR["awq_w4"] # blue
C_FP16_F = "#EAF0F7"
C_QUANT = METHOD_COLOR["gptq_w4"] # red
C_QUANT_F = "#FCEDEB"
C_NEUTRAL = "#555555"
C_NEUTRAL_F = "#F3F3F3"
C_OUT = METHOD_COLOR["bnb_nf4_w4"] # green
C_OUT_F = "#E8F4EE"
C_GOOD = "#5BA56F" # correct-step dot fill
C_BAD = "#D85C58" # failed-step dot fill
# -------- Stage bands --------------------------------------------------
# Each band gets a numbered pill at the top instead of a bare "1." text;
# the pill reads as a stage badge and gives the diagram clearer rhythm.
stage_bands = [
( 0, 27, "1", "Generation", "#F5F8FC", C_FP16),
(29, 56, "2", "Diagnosis", "#F7F7F7", C_NEUTRAL),
(58, 76, "3", "Silver-bullet", "#FDF7F0", "#D77C1F"),
(78, 100, "4", "Restoration", "#EFF6EF", C_OUT),
]
# Approximate width per character for label sizing so we can centre
# the (badge + gap + label) group on each band.
LABEL_FS = 9.5
CHAR_W = LABEL_FS * 0.085 # rough conversion from pt to data units
BADGE_W = 4.0
BADGE_GAP = 1.6 # space between pill and label
for x0, x1, num, label, fc, badge_c in stage_bands:
ax.add_patch(Rectangle((x0, 4), x1 - x0, 64, linewidth=0,
facecolor=fc, zorder=0))
# Stage badge: a coloured pill (number) + stage name. Centre the
# whole group on the band so the number sits to the left of the
# name with a clear gap, no overlap.
cx = (x0 + x1) / 2
label_w = len(label) * CHAR_W
group_w = BADGE_W + BADGE_GAP + label_w
pill_left = cx - group_w / 2
ax.add_patch(FancyBboxPatch((pill_left, 69.5), BADGE_W, 3.6,
boxstyle="round,pad=0.05,rounding_size=1.4",
linewidth=0, facecolor=badge_c, zorder=2))
ax.text(pill_left + BADGE_W / 2, 71.3, num,
ha="center", va="center",
fontsize=8.5, color="white", fontweight="bold", zorder=3)
ax.text(pill_left + BADGE_W + BADGE_GAP, 71.3, label,
ha="left", va="center",
fontsize=LABEL_FS, color="#222", fontweight="bold", zorder=3)
# -------- Helpers ------------------------------------------------------
TITLE_OFFSET = 3.0
SUB_TOP_GAP = 7.5 # raised to clear two-line titles like "Restored\nquant. model"
LINE_SPACING = 2.8
def box(x, y, w, h, label, fc=C_NEUTRAL_F, ec=C_NEUTRAL,
label_fs=8.5, sub_fs=6.8, sub_lines=None,
label_y_frac=None):
"""Rounded box with top-anchored text. If label_y_frac is given the
title sits at that fraction of the box height (used when there's a
mini-visual taking up the lower portion of the box)."""
p = FancyBboxPatch((x, y), w, h,
boxstyle="round,pad=0.02,rounding_size=0.7",
linewidth=1.0, edgecolor=ec, facecolor=fc, zorder=2)
ax.add_patch(p)
cx = x + w / 2
y_top = y + h
if label_y_frac is not None:
ax.text(cx, y + h * label_y_frac, label,
ha="center", va="center",
fontsize=label_fs, color="#1A1A1A", zorder=3)
elif sub_lines:
ax.text(cx, y_top - TITLE_OFFSET, label, ha="center", va="center",
fontsize=label_fs, color="#1A1A1A", zorder=3)
for i, line in enumerate(sub_lines):
ax.text(cx, y_top - SUB_TOP_GAP - i * LINE_SPACING, line,
ha="center", va="center",
fontsize=sub_fs, color=GREY_TEXT, zorder=3)
else:
ax.text(cx, y + h / 2, label, ha="center", va="center",
fontsize=label_fs, color="#1A1A1A", zorder=3)
return (x, y, w, h)
def arrow(src_xy, dst_xy, label=None, rad=0.0, offset=0,
lw=0.9, color="#444444"):
a = FancyArrowPatch(src_xy, dst_xy,
arrowstyle="->,head_width=3,head_length=4",
connectionstyle=f"arc3,rad={rad}",
linewidth=lw, color=color, zorder=1,
shrinkA=2, shrinkB=2)
ax.add_patch(a)
if label:
mx = (src_xy[0] + dst_xy[0]) / 2
my = (src_xy[1] + dst_xy[1]) / 2 + offset
ax.text(mx, my, label, ha="center", va="center",
fontsize=6.8, color="#333",
bbox=dict(boxstyle="round,pad=0.18",
fc="white", ec="none", alpha=0.85),
zorder=3)
def step_dots(cx, cy, n, fail_idx=None, r=0.85, gap=2.4,
ok_color=C_GOOD, fail_color=C_BAD,
ok_alpha=0.9):
"""A horizontal ribbon of n step-dots centred on (cx, cy). Indices
in `fail_idx` are drawn in fail_color (with a slightly larger ring)
to mark erroneous steps."""
fail_idx = set(fail_idx or [])
total_w = (n - 1) * gap
x0 = cx - total_w / 2
for i in range(n):
x = x0 + i * gap
if i in fail_idx:
ax.add_patch(Circle((x, cy), r * 1.15,
facecolor=fail_color, edgecolor="white",
linewidth=0.7, zorder=4))
else:
ax.add_patch(Circle((x, cy), r,
facecolor=ok_color, edgecolor="white",
linewidth=0.6, alpha=ok_alpha, zorder=4))
# ------------------------------------------------------------------
# Shared row geometry. Every stage's "top" content is anchored to
# TOP_Y_TOP/TOP_Y_BOT and every "bottom" content to BOT_Y_TOP/BOT_Y_BOT
# so the diagram has one clean horizontal rhythm instead of three.
# ------------------------------------------------------------------
TOP_Y_BOT = 44 # bottom of the top row
TOP_Y_TOP = 60 # top of the top row (height = 16)
BOT_Y_BOT = 12 # bottom of the bottom row
BOT_Y_TOP = 36 # top of the bottom row (height = 24, taller for richer content)
ROW_H_TOP = TOP_Y_TOP - TOP_Y_BOT
ROW_H_BOT = BOT_Y_TOP - BOT_Y_BOT
# -------- Stage 1: Generation -----------------------------------------
# FP16 reference path (top): model → CoT-with-blue-dots
model_h = 11
b_fp16_m = box(3, TOP_Y_BOT + (ROW_H_TOP - model_h) / 2, 9, model_h,
"FP16\nmodel",
fc=C_FP16_F, ec=C_FP16, label_y_frac=0.5)
b_fp16_t = box(14, TOP_Y_BOT, 12, ROW_H_TOP, "Reference CoT",
fc=C_FP16_F, ec=C_FP16, label_y_frac=0.80)
step_dots(b_fp16_t[0] + b_fp16_t[2] / 2, b_fp16_t[1] + 5.0, n=5,
ok_color=C_FP16, ok_alpha=0.85)
ax.text(b_fp16_t[0] + b_fp16_t[2] / 2, b_fp16_t[1] + 2.0,
"all correct", ha="center", va="center",
fontsize=6.3, color=GREY_TEXT, style="italic")
arrow((b_fp16_m[0] + b_fp16_m[2], b_fp16_m[1] + b_fp16_m[3] / 2),
(b_fp16_t[0], b_fp16_t[1] + b_fp16_t[3] / 2))
# Quantized candidate path (bottom)
b_q_m = box(3, BOT_Y_BOT + (ROW_H_BOT - model_h) / 2, 9, model_h,
"Quantized\nmodel",
fc=C_QUANT_F, ec=C_QUANT, label_y_frac=0.5)
b_q_t = box(14, BOT_Y_BOT + (ROW_H_BOT - 16) / 2, 12, 16,
"Candidate CoT",
fc=C_QUANT_F, ec=C_QUANT, label_y_frac=0.80)
# Failed dots get a contrasting white outline to read clearly even when
# printed in greyscale or at small size.
step_dots(b_q_t[0] + b_q_t[2] / 2, b_q_t[1] + 5.5, n=5,
fail_idx={2, 4}, ok_color=C_QUANT, ok_alpha=0.40,
fail_color=C_BAD)
ax.text(b_q_t[0] + b_q_t[2] / 2, b_q_t[1] + 2.5,
"AWQ / GPTQ / BnB", ha="center", va="center",
fontsize=6.3, color=GREY_TEXT, style="italic")
arrow((b_q_m[0] + b_q_m[2], b_q_m[1] + b_q_m[3] / 2),
(b_q_t[0], b_q_t[1] + b_q_t[3] / 2))
# -------- Stage 2: Diagnosis ------------------------------------------
# DTW alignment box: top row, two rows of dots with pairing lines.
b_align = box(31, TOP_Y_BOT, 24, ROW_H_TOP, "DTW step alignment",
fc=C_NEUTRAL_F, ec=C_NEUTRAL, label_y_frac=0.82)
ax_cx = b_align[0] + b_align[2] / 2
top_y = b_align[1] + 7.5
bot_y = b_align[1] + 3.0
ref_xs = [ax_cx - 8 + i * 4 for i in range(5)]
cand_xs = [ax_cx - 8 + i * 4 for i in range(5)]
for x in ref_xs:
ax.add_patch(Circle((x, top_y), 0.75, facecolor=C_FP16,
edgecolor="white", linewidth=0.5, zorder=4))
for i, x in enumerate(cand_xs):
col = C_BAD if i in (2, 4) else C_QUANT
ax.add_patch(Circle((x, bot_y), 0.85, facecolor=col,
edgecolor="white", linewidth=0.5, zorder=4))
# Pairing lines. One non-monotone pairing (ref-3 ↔ cand-2) demonstrates
# warping; that's the whole reason DTW is used over index matching.
pairings = [(0, 0), (1, 1), (2, 2), (3, 2), (4, 3), (4, 4)]
for r_i, c_i in pairings:
a = FancyArrowPatch((ref_xs[r_i], top_y - 0.7),
(cand_xs[c_i], bot_y + 0.7),
arrowstyle="-",
connectionstyle="arc3,rad=0.0",
linewidth=0.5, color="#888888", zorder=3,
shrinkA=0, shrinkB=0)
ax.add_patch(a)
# Per-step scoring box: bottom row, taller. Layout from top to bottom:
# 1) title
# 2) "is_correct + error_type" subtitle
# 3) horizontal swatch+label legend (label below swatch — conventional)
# 4) the FFS / ECR / SSR metric pills, anchored INSIDE the box (was
# floating below before)
b_score = box(31, BOT_Y_BOT, 24, ROW_H_BOT, "Per-step scoring",
fc=C_NEUTRAL_F, ec=C_NEUTRAL, label_y_frac=0.88)
sw_cx = b_score[0] + b_score[2] / 2
ax.text(sw_cx, b_score[1] + ROW_H_BOT * 0.74,
"is_correct + error_type", ha="center", va="center",
fontsize=6.8, color=GREY_TEXT, style="italic")
# Swatch row (chip on top, label below).
swatch_y = b_score[1] + ROW_H_BOT * 0.52
swatch_label_y = b_score[1] + ROW_H_BOT * 0.40
et_labels = ["concept.", "method.", "execut.", "logical"]
sw_step = 5.4
for i, et in enumerate(ERROR_TYPES):
sx = sw_cx + (i - 1.5) * sw_step
ax.add_patch(FancyBboxPatch((sx - 1.6, swatch_y - 1.1), 3.2, 2.2,
boxstyle="round,pad=0.02,rounding_size=0.6",
linewidth=0, facecolor=ERROR_COLORS[et],
zorder=4))
ax.text(sx, swatch_label_y, et_labels[i],
ha="center", va="center",
fontsize=6.0, color="#333", zorder=4)
# Metric pills, ANCHORED inside the scoring box (no longer floating).
# Small "yields →" tag on the left makes the relationship explicit.
metric_y = b_score[1] + ROW_H_BOT * 0.18
metric_h = 3.6
metric_w = 4.8
metric_gap = 1.2
metric_specs = [("FFS", "#4E79A7"),
("ECR", "#E15759"),
("SSR", "#59A14F")]
n_m = len(metric_specs)
total_metric_w = n_m * metric_w + (n_m - 1) * metric_gap
yields_w = 6.5
group_left = sw_cx - (total_metric_w + yields_w) / 2 + yields_w
ax.text(group_left - 1.0, metric_y, "yields →",
ha="right", va="center", fontsize=6.5, color=GREY_TEXT,
style="italic", zorder=4)
for i, (name, col) in enumerate(metric_specs):
sx = group_left + i * (metric_w + metric_gap) + metric_w / 2
ax.add_patch(FancyBboxPatch((sx - metric_w / 2, metric_y - metric_h / 2),
metric_w, metric_h,
boxstyle="round,pad=0.05,rounding_size=1.4",
linewidth=0, facecolor=col, alpha=0.92,
zorder=4))
ax.text(sx, metric_y, name, ha="center", va="center",
fontsize=7.5, color="white", fontweight="bold", zorder=5)
# Stage 1 → Stage 2 arrows.
arrow((b_fp16_t[0] + b_fp16_t[2], b_fp16_t[1] + b_fp16_t[3] / 2),
(b_align[0], b_align[1] + b_align[3] * 0.55),
label="ref.", offset=1.5)
arrow((b_q_t[0] + b_q_t[2], b_q_t[1] + b_q_t[3] / 2),
(b_score[0], b_score[1] + b_score[3] * 0.65),
label="cand.", offset=-1.5)
# Internal alignment → scoring arrow (carries DTW pairs to the scorer).
arrow((b_align[0] + b_align[2] / 2, b_align[1]),
(b_score[0] + b_score[2] / 2, b_score[1] + b_score[3]),
rad=0.0, label="pairs", offset=0)
# -------- Stage 3: Silver-bullet construction -------------------------
# A compact box centred in the band (between the two rows) — previous
# version spanned the full band height which left ~28 units of empty
# space. The 4-chip strip honestly conveys "stratified by error type"
# without implying equal-quartile sampling (the donut+"4×" did).
b_filter_x, b_filter_w = 60, 14
b_filter_h = 30
band_mid_y = (BOT_Y_BOT + TOP_Y_TOP) / 2
b_filter_y = band_mid_y - b_filter_h / 2
box(b_filter_x, b_filter_y, b_filter_w, b_filter_h, "Silver-bullet\ndataset",
fc="#FFF2E5", ec="#D77C1F", label_y_frac=0.83)
chip_y = b_filter_y + b_filter_h * 0.48
chip_w = 2.4; chip_h = 2.4; chip_gap = 0.8
chip_total_w = 4 * chip_w + 3 * chip_gap
chip_left = b_filter_x + b_filter_w / 2 - chip_total_w / 2
for i, et in enumerate(ERROR_TYPES):
ax.add_patch(FancyBboxPatch((chip_left + i * (chip_w + chip_gap),
chip_y - chip_h / 2),
chip_w, chip_h,
boxstyle="round,pad=0.02,rounding_size=0.6",
linewidth=0,
facecolor=ERROR_COLORS[et], zorder=4))
ax.text(b_filter_x + b_filter_w / 2, chip_y + chip_h / 2 + 1.8,
"stratified by", ha="center", va="center",
fontsize=6.4, color=GREY_TEXT, style="italic")
ax.text(b_filter_x + b_filter_w / 2, chip_y - chip_h / 2 - 1.8,
"error type", ha="center", va="center",
fontsize=6.4, color=GREY_TEXT, style="italic")
ax.text(b_filter_x + b_filter_w / 2, b_filter_y + b_filter_h * 0.13,
"$N\\!\\leq\\!500$ failed problems",
ha="center", va="center",
fontsize=6.5, color=GREY_TEXT)
# Stage 2 → Stage 3: scoring box (bottom row) → silver-bullet (centre).
# Source from the scoring box's upper-right; target the box's left
# edge mid-height. Keeps the arrow in clear whitespace above the chips.
arrow((b_score[0] + b_score[2], b_score[1] + b_score[3] * 0.78),
(b_filter_x, b_filter_y + b_filter_h * 0.50),
label="scored\nsteps", offset=0)
# -------- Stage 4: Restoration ----------------------------------------
# Top-row QLoRA box anchored to the same TOP row as Stage 1/2.
b_qlora = box(79, TOP_Y_BOT, 19, ROW_H_TOP, "QLoRA fine-tune",
sub_lines=["rank 16, $\\alpha\\!=\\!32$",
"on quantized base",
"$<\\!0.4\\%$ trainable"],
fc=C_OUT_F, ec=C_OUT)
# The arrow itself enters the middle of the QLoRA box, but its mid-point
# label is offset upward so it sits between the box top and the first
# sub-line — otherwise the label crashed into the "<0.4% trainable"
# text in the QLoRA box (same y-range, both rendered, looked merged).
arrow((b_filter_x + b_filter_w, b_filter_y + b_filter_h * 0.65),
(b_qlora[0], b_qlora[1] + b_qlora[3] * 0.5),
label="train set", offset=7)
# Bottom-row Restored model box anchored to the same BOT row.
b_out = box(79, BOT_Y_BOT, 19, ROW_H_BOT, "Restored\nquant. model",
sub_lines=["adapter merged,",
"re-quantized"],
fc=C_OUT_F, ec=C_OUT)
arrow((b_qlora[0] + b_qlora[2] / 2, b_qlora[1]),
(b_out[0] + b_out[2] / 2, b_out[1] + b_out[3]),
label="adapter", offset=0)
fig.tight_layout()
fig.savefig(output_path)
plt.close(fig)
print(f" Paper fig 0 saved: {output_path}")
# ---------------------------------------------------------------------------
# Fig 4 — FFS distribution per method (base vs restored overlay)
# ---------------------------------------------------------------------------
def fig_paper_4_ffs_distribution(diagnosis_dir: str, primary_model: str,
primary_benchmark: str, output_path: str):
"""3 small-multiple panels, one per quantization method. Each panel shows
the FFS histogram — base (solid fill) vs restored (outlined) — for
problems that failed. Bin edges chosen to resolve the dominant 0-5 range
and then lump 6+."""
# Slightly taller than before so the legend can sit under the
# shared x-axis labels without colliding with them.
fig, axes = plt.subplots(1, 3, figsize=(6.75, 2.55),
sharey=True, constrained_layout=False)
fig.subplots_adjust(left=0.08, right=0.99, top=0.88, bottom=0.32,
wspace=0.12)
# Wider buckets in the tail — the data is effectively zero beyond step 5,
# so lumping 6+ into a single "6+" bucket avoids the overlapping multi-char
# labels that "6-10 / 11-20 / 21+" produced at this figure width.
bin_edges = np.array([0, 1, 2, 3, 4, 5, 6, 31])
bin_labels = ["0", "1", "2", "3", "4", "5", "6+"]
for ax, method in zip(axes, METHOD_ORDER):
base_traces = _load_diagnosis(diagnosis_dir, primary_model, method, primary_benchmark)
rest_traces = _load_diagnosis(diagnosis_dir, primary_model, method + "_restored", primary_benchmark)
def ffs_vec(traces):
vals = []
for t in traces:
f = _first_failure_step(t.get("steps", []) or [])
if f is not None:
vals.append(f)
return np.array(vals)
base_ffs = ffs_vec(base_traces)
rest_ffs = ffs_vec(rest_traces)
base_h, _ = np.histogram(base_ffs, bins=bin_edges)
rest_h, _ = np.histogram(rest_ffs, bins=bin_edges)
x = np.arange(len(bin_labels))
width = 0.38
color = METHOD_COLOR[method]
color_rest = METHOD_COLOR[method + "_restored"]
ax.bar(x - width / 2, base_h, width=width, color=color,
label="Quantized", edgecolor="white", linewidth=0.6)
ax.bar(x + width / 2, rest_h, width=width, facecolor="none",
edgecolor=color, linewidth=1.0, label="Restored")
ax.set_xticks(x)
ax.set_xticklabels(bin_labels)
ax.yaxis.grid(True)
ax.set_axisbelow(True)
ax.set_xlabel("First failure step")
ax.text(0.98, 0.96, METHOD_PRETTY[method], transform=ax.transAxes,
ha="right", va="top", fontsize=7.5, color=GREY_TEXT)
axes[0].set_ylabel("Failed problems")
# Shared legend — anchored below the x-axis labels (there is now
# explicit bottom padding via subplots_adjust) so it cannot overlap
# the per-panel "First failure step" captions.
handles = [
plt.Rectangle((0, 0), 1, 1, facecolor=GREY_TEXT, edgecolor="white",
linewidth=0.6, label="Quantized"),
plt.Rectangle((0, 0), 1, 1, facecolor="none", edgecolor=GREY_TEXT,
linewidth=1.0, label="Restored"),
]
fig.legend(handles=handles, loc="lower center",
bbox_to_anchor=(0.5, 0.02),
ncol=2, handlelength=1.6, handletextpad=0.5, columnspacing=1.6,
frameon=False)
# Subtitle-style note at top-left
fig.text(0.01, 0.95, f"{primary_model} · {BENCH_PRETTY[primary_benchmark]}",
ha="left", va="bottom", fontsize=7, color=GREY_TEXT)
fig.savefig(output_path)
plt.close(fig)
print(f" Paper fig 4 saved: {output_path}")
# ---------------------------------------------------------------------------
# Fig 5 — Qualitative trace example (FP16 vs quantized, diagnosed)
# ---------------------------------------------------------------------------
def fig_paper_5_trace_example(diagnosis_dir: str, segmented_dir: str,
primary_model: str, primary_benchmark: str,
output_path: str,
method: str = "gptq_w4",
problem_id: str = "math500_12",
n_steps_to_show: int = 4):
"""Side-by-side diagnosed trace: FP16 reference (left) vs quantized
candidate (right), with explicit visual encodings of the diagnosis the
paper claims to produce:
• per-step ✓ / ✗ status badges in the gutter
• a "First Failure Step" callout band on the row of the very first
quantized failure (FFS = the headline metric of the paper)
• subtle horizontal alignment ribbons in the inter-column gap that
echo the DTW pairing depicted in fig 0
• error-type chip embedded in the failing box (same colour palette
as figs 0/2/12, so the figure stays consistent across the paper)
"""
import textwrap
from matplotlib.patches import FancyBboxPatch as _FBPatch
quant_traces = _load_diagnosis(diagnosis_dir, primary_model, method, primary_benchmark)
fp16_traces = _load_segmented_fp16(segmented_dir, primary_model, primary_benchmark)
fp16 = next((t for t in fp16_traces if t.get("problem_id") == problem_id), None)
qnt = next((t for t in quant_traces if t.get("problem_id") == problem_id), None)
if not fp16 or not qnt:
print(f" [SKIP] Missing traces for {problem_id}")
return
# The segmented/diagnosis JSONLs don't carry the question text; pull it
# from the inference output (which has the original prompt).
def _load_question(quant_subdir: str) -> Optional[str]:
path = os.path.join("results/inference", quant_subdir, primary_model,
f"{primary_benchmark}_run0.jsonl")
if not os.path.exists(path):
return None
with open(path) as f:
for line in f:
row = json.loads(line)
if row.get("problem_id") == problem_id:
return row.get("question") or row.get("problem_text")
return None
question_text = (_load_question("fp16") or _load_question(method) or "")
fp16_steps = (fp16.get("steps") or [])[:n_steps_to_show]
qnt_steps = (qnt.get("steps") or [])[:n_steps_to_show]
# First-failure-step index (in the quantized trace). The visual callout
# is anchored on the row that matches this step. None ⇒ no callout.
ffs = _first_failure_step(qnt_steps)
# ------------------------------------------------------------------
# Pre-wrap both columns. Wrap width and line cap are both small enough
# that the inter-column gap doesn't get squeezed (we now reserve a
# 14%-wide gutter for the alignment ribbon + ✓/✗ badges).
# ------------------------------------------------------------------
# WRAP_W tuned for the new 7.4" canvas with 0.18-ratio gutter:
# ~47 chars fit per line at 8pt. MAX_LINES=6 so a typical 240-260 char
# step renders as full prose (no "[…]" marker) — earlier 5-line cap
# produced bogus "[…]" tails on steps whose text would otherwise fit.
WRAP_W = 50
MAX_LINES = 6
def _prewrap(steps):
out = []
for s in steps:
text = _truncate(s.get("text", ""), WRAP_W * MAX_LINES)
lines = textwrap.wrap(text, width=WRAP_W) or [""]
if len(lines) > MAX_LINES:
lines = lines[:MAX_LINES]
# End on whole-word boundary then append the ellipsis as a
# separate visual marker, so it reads as "(continues)" not
# as a typo.
lines[-1] = lines[-1].rstrip() + " […]"
out.append({"lines": lines, "s": s})
return out
wl = _prewrap(fp16_steps)
wr = _prewrap(qnt_steps)
n_rows = max(len(wl), len(wr))
while len(wl) < n_rows: wl.append(None)
while len(wr) < n_rows: wr.append(None)
row_lines = [
max((len(wl[i]["lines"]) if wl[i] else 0),
(len(wr[i]["lines"]) if wr[i] else 0))
for i in range(n_rows)
]
row_units = [lc + 1.0 for lc in row_lines]
total_units = sum(row_units) + (n_rows - 1) * 0.4
LINE_INCH = 0.165
CONTENT_IN = total_units * LINE_INCH
if question_text:
q_lines = max(1, min(4, len(textwrap.wrap(question_text, width=110))))
HEADER_IN = 0.55 + 0.16 * q_lines
else:
HEADER_IN = 0.60
FOOTER_IN = 0.32
fig_h = HEADER_IN + CONTENT_IN + FOOTER_IN
fig_w = 7.4 # +0.4" wider than the previous version to give the
# alignment-ribbon gutter and ✓/✗ badges proper breathing room
fig = plt.figure(figsize=(fig_w, fig_h))
top_frac = 1.0 - HEADER_IN / fig_h
bottom_frac = FOOTER_IN / fig_h
# Three-column gridspec: [left text] [centre gutter] [right text].
# The centre gutter axis hosts ✓/✗ status badges, the alignment ribbon
# connecting paired rows, and the FFS callout marker — all live in the
# same coordinate space so they line up perfectly with the rows.
gs = fig.add_gridspec(1, 3,
width_ratios=[1.0, 0.18, 1.0],
wspace=0.0,
left=0.02, right=0.98,
top=top_frac, bottom=bottom_frac)
ax_l = fig.add_subplot(gs[0, 0]); ax_l.set_axis_off()
ax_g = fig.add_subplot(gs[0, 1]); ax_g.set_axis_off()
ax_r = fig.add_subplot(gs[0, 2]); ax_r.set_axis_off()
ax_g.set_xlim(0, 1); ax_g.set_ylim(0, 1)
# ----- Problem header -------------------------------------------------
if question_text:
problem_wrap = textwrap.fill(_truncate(question_text, 320), width=110)
fig.text(0.5, 1.0 - 0.10 / fig_h, problem_wrap,
ha="center", va="top", fontsize=7.5, color="#222222",
bbox=dict(boxstyle="round,pad=0.5", fc="#F4F6F9",
ec="#B7C1CE", lw=0.7))
# ----- Column headers -------------------------------------------------
ax_l.text(0.5, 1.015, "FP16 reference", transform=ax_l.transAxes,
ha="center", va="bottom", fontsize=9,
color=METHOD_COLOR["awq_w4"], fontweight="bold")
ax_r.text(0.5, 1.015, f"{METHOD_PRETTY[method]} (quantized)",
transform=ax_r.transAxes, ha="center", va="bottom",
fontsize=9, color=METHOD_COLOR[method], fontweight="bold")
# Gutter sub-header — spelled out instead of "diag." abbreviation, so
# the figure stays self-explanatory to a reader who hasn't read §3 yet.
ax_g.text(0.5, 1.015, "verdict", transform=ax_g.transAxes,
ha="center", va="bottom", fontsize=7.5,
color=GREY_TEXT, style="italic")
# ----- Shared y-grid --------------------------------------------------
y_top_axes = 0.985
y_bot_axes = 0.015
avail = y_top_axes - y_bot_axes
unit = avail / max(total_units, 1)
row_y_top = []
row_y_center = []
row_y_bot = []
y_cursor = y_top_axes
for i in range(n_rows):
block_h = row_units[i] * unit
y_top = y_cursor
y_cursor -= block_h
y_center = (y_top + y_cursor) / 2 + unit * 0.20
row_y_top.append(y_top)
row_y_center.append(y_center)
row_y_bot.append(y_cursor)
if i < n_rows - 1:
y_cursor -= 0.4 * unit
# ------------------------------------------------------------------
# Render each column. Step badges (the small numbered circle) sit in
# the body axes' left margin; ✓ / ✗ status icons sit in the GUTTER
# axis so they read as a per-row diagnosis verdict, independent of
# either text column.
# ------------------------------------------------------------------
def render(ax, wrapped_rows, accent_color, is_quant):
for i, w in enumerate(wrapped_rows):
if w is None:
continue
y_center = row_y_center[i]
is_correct = w["s"].get("is_correct", True)
err_type = w["s"].get("error_type") or ""
is_fail = is_quant and (is_correct is False)
if is_fail:
ec = ERROR_COLORS.get(err_type, "#E76F51")
fc = "#FFF7F4"
lw = 1.2
else:
ec = "#D6D9DD"
fc = "#FAFBFC" if is_quant else "#F6FAF7"
lw = 0.5
# Numbered step badge in the column's left margin.
ax.text(0.005, y_center, f"{i}",
transform=ax.transAxes, va="center", ha="left",
fontsize=8.5, color=accent_color, fontweight="bold",
bbox=dict(boxstyle="circle,pad=0.25",
fc="white", ec=accent_color, lw=0.9))
# Step body.
body = "\n".join(w["lines"])
ax.text(0.07, y_center, body,
transform=ax.transAxes, va="center", ha="left",
fontsize=8, color="#222222",
bbox=dict(boxstyle="round,pad=0.5",
fc=fc, ec=ec, lw=lw))
# Error-type chip inline at the bottom-right of the failing box.
if is_fail and err_type:
ax.text(0.985, y_center - (row_lines[i] * 0.5 - 0.5) * unit,
err_type,
transform=ax.transAxes, va="top", ha="right",
fontsize=6.5, color="white",
bbox=dict(boxstyle="round,pad=0.28",
fc=ERROR_COLORS.get(err_type, "#E76F51"),
ec="none"))
render(ax_l, wl, METHOD_COLOR["awq_w4"], is_quant=False)
render(ax_r, wr, METHOD_COLOR[method], is_quant=True)
# ------------------------------------------------------------------
# Gutter visualisations: alignment ribbon + ✓ / ✗ status icons.
# ------------------------------------------------------------------
OK_COLOR = "#3CA56F"
FAIL_COLOR = "#D85C58"
for i in range(n_rows):
if wl[i] is None or wr[i] is None:
continue
yc = row_y_center[i]
# Thin alignment ribbon spanning the gutter — echoes DTW pairing.
ax_g.plot([0.05, 0.95], [yc, yc], color="#C6CCD3",
linewidth=0.6, zorder=1)
# Per-row verdict from the quantized side (the FP16 trace is taken
# as ground truth in this figure, so its ✓ is implicit).
is_correct_q = wr[i]["s"].get("is_correct", True)
is_fail_q = is_correct_q is False
glyph = "✗" if is_fail_q else "✓"
gcol = FAIL_COLOR if is_fail_q else OK_COLOR
# White-ringed pill for the verdict, sized to read at a glance.
ax_g.text(0.5, yc, glyph, transform=ax_g.transAxes,
ha="center", va="center",
fontsize=8.5, color="white", fontweight="bold",
bbox=dict(boxstyle="circle,pad=0.30",
fc=gcol, ec="white", lw=1.0),
zorder=4)
# ------------------------------------------------------------------
# First Failure Step callout. Anchored to the FFS row, sits in the
# left-most margin of the figure (left of the FP16 column) and points
# right toward the failing row. This is the central concept of the
# paper, so we make it visually unmissable without disrupting layout.
# ------------------------------------------------------------------
if ffs is not None and 0 <= ffs < n_rows:
yc = row_y_center[ffs]
from matplotlib.patches import Rectangle as _Rect
# Yellow-tinted band the full row width. Slightly stronger fill
# (alpha 0.75 + thin amber edge) than the previous 0.55-alpha tint,
# which read too faintly in print.
band_h_axes = max(row_lines[ffs] * 0.55 * unit, 0.030)
band_y_fig = bottom_frac + (yc - band_h_axes / 2) * (top_frac - bottom_frac)
band_h_fig = band_h_axes * (top_frac - bottom_frac)
fig.add_artist(_Rect((0.02, band_y_fig), 0.96, band_h_fig,
facecolor="#FFF1B5", edgecolor="#E5C46A",
linewidth=0.5, alpha=0.75, zorder=0))
# FFS pill in the gutter axis itself (NOT the figure margin), so
# it sits cleanly above the verdict pill of that row, doesn't
# collide with the FP16 column's left edge, and inherits the
# gutter's coordinate space so it tracks row layout exactly.
ax_g.text(0.5, yc + band_h_axes * 0.55,
"first failure step",
transform=ax_g.transAxes,
ha="center", va="bottom",
fontsize=6.5, color="#7A5A0F", fontweight="bold",
bbox=dict(boxstyle="round,pad=0.30",
fc="#FFE066", ec="#D4A516", lw=0.6),
zorder=5)
# ----- Footer caption -------------------------------------------------
fig.text(0.02, FOOTER_IN / fig_h * 0.4,
f"Problem: {problem_id} · Model: {primary_model} · "
f"Method: {METHOD_PRETTY[method]}",
ha="left", va="center", fontsize=6.8, color=GREY_TEXT)
# Inline legend at the right side of the footer. Removed the FFS
# entry because the visible "first failure step" yellow pill already
# serves as its own legend, and listing a "░" pattern character that
# didn't match the actual yellow-band marker was actively misleading.
fig.text(0.98, FOOTER_IN / fig_h * 0.4,
"✓ correct step ✗ failed step",
ha="right", va="center", fontsize=6.8, color=GREY_TEXT)
fig.savefig(output_path)
plt.close(fig)
print(f" Paper fig 5 saved: {output_path}")
def _truncate(s: str, n: int) -> str:
s = (s or "").replace("\n", " ").strip()
return s if len(s) <= n else s[:n - 1].rstrip() + "…"
# ---------------------------------------------------------------------------
# Fig 11 — Correlation between the novel metrics
# ---------------------------------------------------------------------------
def fig_paper_11_metric_correlation(data, output_path: str):
"""Pearson correlation between (Accuracy, FFS, ECR, avg_token_count)
computed across all (model, benchmark, quant) cells. If all four were
redundant the matrix would be near-perfect 1s; non-trivial off-diagonals
justify the 3-metric framework as a strict extension of accuracy alone.
"""
fields = ["accuracy", "avg_ffs", "median_ffs", "ecr"]
labels = ["Accuracy", "Avg FFS", "Median FFS", "ECR"]
rows = []
for (model, quant, bench), entry in data.items():
row = [entry.get(f) for f in fields]
if any(v is None or (isinstance(v, float) and (v != v)) for v in row):
continue
rows.append(row)
if not rows:
print(" [SKIP] fig 11: no complete rows to correlate")
return
arr = np.array(rows, dtype=float)
# Guard against fields that are all identical / all NaN (would nan out corr)
corr = np.corrcoef(arr.T)
fig, ax = plt.subplots(figsize=(3.5, 3.0), constrained_layout=True)
im = ax.imshow(corr, cmap="RdBu_r", vmin=-1, vmax=1, aspect="auto")
for i in range(len(fields)):
for j in range(len(fields)):
c = "white" if abs(corr[i, j]) > 0.6 else "#222"
ax.text(j, i, f"{corr[i, j]:.2f}", ha="center", va="center",
fontsize=8, color=c)
ax.set_xticks(range(len(fields)))
ax.set_yticks(range(len(fields)))
ax.set_xticklabels(labels, rotation=20, ha="right")
ax.set_yticklabels(labels)
cbar = plt.colorbar(im, ax=ax, fraction=0.046, pad=0.04,
ticks=[-1, -0.5, 0, 0.5, 1])
cbar.ax.tick_params(labelsize=7)
cbar.set_label("Pearson r", fontsize=8)
ax.text(1.0, 1.05,
f"N = {len(rows)} cells (model × benchmark × quant)",
transform=ax.transAxes, ha="right", va="bottom",
fontsize=7, color=GREY_TEXT)
fig.savefig(output_path)
plt.close(fig)
print(f" Paper fig 11 saved: {output_path}")
# ---------------------------------------------------------------------------
# Fig 12 — Error-type × step-depth heatmap
# ---------------------------------------------------------------------------
def fig_paper_12_error_x_depth(diagnosis_dir: str, primary_model: str,
primary_benchmark: str, output_path: str,
max_depth: int = 15):
"""For the primary (model, benchmark), pool failed steps from all three
base quantization methods and build a heatmap P(error_type | depth).
Reads: for each quant method's diagnosis jsonl, every step with
`is_correct == False` gets counted at its `index` under its `error_type`.
Rows are error types, columns are step depths 0..max_depth-1. Cells are
fraction of errors at that depth that have this type (columns sum to 1).
Per-step error-type labels: this figure is rendered from the LLM-judge
re-classification (results/diagnosis_llm/) when available, matching the
aggregate distribution reported in §5.4 and the convention announced in
§3.5. Falls back to the rule-based diagnosis directory only if the
LLM-judge directory is missing for the primary cell.
"""
counts = np.zeros((len(ERROR_TYPES), max_depth), dtype=float)
# Prefer LLM-judge labels for the primary cell; fall back to rule-based
# diagnosis directory if no LLM-judge re-classification was run.
llm_root = os.path.join(os.path.dirname(diagnosis_dir.rstrip("/")),
"diagnosis_llm")
primary_dir = llm_root if os.path.isdir(llm_root) else diagnosis_dir
for method in METHOD_ORDER:
traces = _load_diagnosis(primary_dir, primary_model, method, primary_benchmark)
for t in traces:
for s in (t.get("steps") or []):
if s.get("is_correct") is not False:
continue
d = s.get("index", 0)
et = s.get("error_type")
if et in ERROR_TYPES and 0 <= d < max_depth:
counts[ERROR_TYPES.index(et), d] += 1
col_sums = counts.sum(axis=0, keepdims=True)
with np.errstate(invalid="ignore", divide="ignore"):
prop = np.where(col_sums > 0, counts / col_sums, np.nan)
fig, ax = plt.subplots(figsize=(6.75, 2.3), constrained_layout=True)
im = ax.imshow(prop, cmap="YlOrRd", aspect="auto", vmin=0, vmax=1)
ax.set_xticks(range(max_depth))
ax.set_xticklabels(range(max_depth))
ax.set_yticks(range(len(ERROR_TYPES)))
ax.set_yticklabels([e.capitalize() for e in ERROR_TYPES])
ax.set_xlabel("Step depth")
for i in range(len(ERROR_TYPES)):
for j in range(max_depth):
v = prop[i, j]
if np.isnan(v):
continue
c = "white" if v > 0.5 else "#222"
ax.text(j, i, f"{v:.0%}" if v >= 0.05 else "",
ha="center", va="center", fontsize=6.5, color=c)
# Per-column total count, annotated on top (how many errors are at that depth).
col_totals = counts.sum(axis=0).astype(int)
for j, n in enumerate(col_totals):
ax.text(j, -0.8, str(n), ha="center", va="center", fontsize=6.3, color=GREY_TEXT)
ax.text(-0.6, -0.8, "$n$=", ha="right", va="center", fontsize=6.3, color=GREY_TEXT)
cbar = plt.colorbar(im, ax=ax, fraction=0.035, pad=0.02,
ticks=[0, 0.25, 0.5, 0.75, 1])
cbar.ax.tick_params(labelsize=7)
cbar.set_label("Fraction of errors at that depth", fontsize=7.5)
# Metadata tag on the colorbar axis (empty space there) so it doesn't
# collide with the "n=" counts row sitting at y = -0.8 inside this axes.
cbar.ax.text(0.5, 1.06,
f"{primary_model} · {BENCH_PRETTY[primary_benchmark]}",
transform=cbar.ax.transAxes, ha="center", va="bottom",
fontsize=7.5, color=GREY_TEXT)
fig.savefig(output_path)
plt.close(fig)
print(f" Paper fig 12 saved: {output_path}")
# ---------------------------------------------------------------------------
# Fig 13 — Error-type reduction after restoration (slopegraph)
# ---------------------------------------------------------------------------
def fig_paper_13_error_reduction(data, primary_model: str,
primary_benchmark: str, output_path: str):
"""Slopegraph: for each (method, error_type), draw a line from the
pre-restoration error rate (per 100 problems) to the post-restoration
rate. Lines that slope down = restoration reduced that error type;
lines that slope up = restoration made it worse. Side-by-side panels
per quantization method keep the reader from confusing families.
"""
fig, axes = plt.subplots(1, len(METHOD_ORDER),
figsize=(6.75, 2.6), sharey=True,
constrained_layout=True)
if len(METHOD_ORDER) == 1:
axes = [axes]
for ax, method in zip(axes, METHOD_ORDER):
base = data.get((primary_model, method, primary_benchmark)) or {}
rest = data.get((primary_model, method + "_restored", primary_benchmark)) or {}
base_acc = base.get("accuracy") or 0.0
rest_acc = rest.get("accuracy") or 0.0
base_dist = base.get("error_type_dist") or {}
rest_dist = rest.get("error_type_dist") or {}
# Errors per 100 problems of each type, before vs after
before = {e: (1 - base_acc) * 100 * base_dist.get(e, 0) for e in ERROR_TYPES}
after = {e: (1 - rest_acc) * 100 * rest_dist.get(e, 0) for e in ERROR_TYPES}
y_max = max(list(before.values()) + list(after.values()) + [1])
# Draw each error type's line
for e in ERROR_TYPES:
b, a = before[e], after[e]
color = ERROR_COLORS[e]
ax.plot([0, 1], [b, a], "-", color=color, linewidth=1.6, alpha=0.9, zorder=2)
ax.scatter([0], [b], s=28, color=color, edgecolor="white",
linewidth=0.8, zorder=3)
ax.scatter([1], [a], s=28, color=color, edgecolor="white",
linewidth=0.8, zorder=3)
# Annotate on the restored side
ax.annotate(f"{e[:5]} {a:.1f}", xy=(1, a), xytext=(4, 0),
textcoords="offset points", ha="left", va="center",
fontsize=6.5, color=color)
ax.set_xticks([0, 1])
ax.set_xticklabels(["Quantized", "Restored"])
ax.set_xlim(-0.12, 1.5)
ax.set_ylim(0, y_max * 1.15 + 1)
ax.text(0.98, 0.97, METHOD_PRETTY[method], transform=ax.transAxes,
ha="right", va="top", fontsize=7.5, color=GREY_TEXT)
ax.yaxis.grid(True); ax.set_axisbelow(True)
axes[0].set_ylabel("Errors per 100 problems")
fig.savefig(output_path)
plt.close(fig)
print(f" Paper fig 13 saved: {output_path}")
# ---------------------------------------------------------------------------
# Entrypoint
# ---------------------------------------------------------------------------
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--metrics", required=True)
parser.add_argument("--diagnosis", default="results/diagnosis",
help="Diagnosis directory (for FFS dist. and trace example)")
parser.add_argument("--segmented", default="results/segmented",
help="Segmented directory (for FP16 reference in trace example)")
parser.add_argument("--output", required=True)
parser.add_argument("--primary-model", default="r1-qwen-7b")
parser.add_argument("--primary-benchmark", default="math500")
parser.add_argument("--trace-method", default="gptq_w4",
help="Quant method featured in the qualitative trace figure")
parser.add_argument("--trace-problem", default="math500_12",
help="Problem id featured in the qualitative trace figure")
parser.add_argument("--models", nargs="*", default=None)
args = parser.parse_args()
os.makedirs(args.output, exist_ok=True)
data = load_all(args.metrics)
if not data:
print(f"No metrics found in {args.metrics}")
return
all_models = sorted({m for (m, _, _) in data.keys()})
models = args.models if args.models else all_models
print(f"Primary model: {args.primary_model} Models: {models}")
fig_paper_0_pipeline(os.path.join(args.output, "fig_paper_0_pipeline.pdf"))
fig_paper_1_ssr(data, args.primary_model,
os.path.join(args.output, "fig_paper_1_ssr.pdf"))
fig_paper_2_error_mix(data, args.primary_model, args.primary_benchmark,
os.path.join(args.output, "fig_paper_2_error_mix.pdf"))
fig_paper_3_forest(data, models,
os.path.join(args.output, "fig_paper_3_forest.pdf"))
fig_paper_4_ffs_distribution(args.diagnosis, args.primary_model,
args.primary_benchmark,
os.path.join(args.output, "fig_paper_4_ffs_dist.pdf"))
fig_paper_5_trace_example(args.diagnosis, args.segmented,
args.primary_model, args.primary_benchmark,
os.path.join(args.output, "fig_paper_5_trace_example.pdf"),
method=args.trace_method,
problem_id=args.trace_problem)
fig_paper_11_metric_correlation(data,
os.path.join(args.output, "fig_paper_11_metric_correlation.pdf"))
fig_paper_12_error_x_depth(args.diagnosis, args.primary_model,
args.primary_benchmark,
os.path.join(args.output, "fig_paper_12_error_x_depth.pdf"))
fig_paper_13_error_reduction(data, args.primary_model,
args.primary_benchmark,
os.path.join(args.output, "fig_paper_13_error_reduction.pdf"))
table_paper_tex(data, models,
os.path.join(args.output, "table_paper.tex"))
print(f"\nPaper artifacts written under {args.output}")
if __name__ == "__main__":
main()
|