File size: 19,862 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 | """
StepProbe: Paper Figure Generation
Generates all figures for the paper:
1. Step Survival Rate (SSR) curves
2. Error Type Distribution heatmap
3. First Failure Step (FFS) distributions
4. Accuracy vs. bit-width degradation
5. Error Cascade Rate by model size
6. Restoration before/after comparison
"""
import json
import os
import sys
import glob
from typing import List, Dict
import numpy as np
import matplotlib.pyplot as plt
import matplotlib
matplotlib.rcParams.update({
"font.family": "sans-serif",
"font.size": 11,
"axes.titlesize": 13,
"axes.labelsize": 12,
"xtick.labelsize": 10,
"ytick.labelsize": 10,
"legend.fontsize": 10,
"figure.dpi": 150,
"savefig.dpi": 300,
"savefig.bbox": "tight",
})
sys.path.insert(0, os.path.dirname(os.path.dirname(os.path.abspath(__file__))))
from stepprobe.utils import load_json
# Color palette (colorblind-friendly)
# Paired palette: the "restored" variant shares a hue with its base so the
# eye can pair them instantly. Base = saturated, restored = desaturated.
COLORS = {
"fp16": "#1F4E79", # navy — reference
"awq_w4": "#A23B72",
"awq_w4_restored": "#D9A9C4",
"gptq_w4": "#C73E1D",
"gptq_w4_restored": "#EFB4A6",
"bnb_nf4_w4": "#0B7A75",
"bnb_nf4_w4_restored": "#8FC8C5",
}
ERROR_COLORS = {
"conceptual": "#2E86AB",
"methodological": "#A23B72",
"executional": "#F18F01",
"logical": "#C73E1D",
}
# Canonical ordering so paired (base, restored) bars always sit next to each
# other and colors line up across figures.
QUANT_ORDER = [
"awq_w4", "awq_w4_restored",
"gptq_w4", "gptq_w4_restored",
"bnb_nf4_w4", "bnb_nf4_w4_restored",
]
def _pretty_label(q: str) -> str:
"""Short, lowercase-friendly label: 'gptq_w4_restored' -> 'GPTQ w4 (restored)'."""
if q.endswith("_restored"):
base = q[: -len("_restored")]
return f"{_pretty_label(base)} (restored)"
if q == "bnb_nf4_w4":
return "BnB NF4"
if q == "fp16":
return "FP16"
parts = q.split("_w")
if len(parts) == 2 and parts[1].isdigit():
return f"{parts[0].upper()} w{parts[1]}"
return q
def _sorted_results(results):
"""Return results ordered by QUANT_ORDER so figures are consistent."""
order = {q: i for i, q in enumerate(QUANT_ORDER)}
return sorted(results, key=lambda r: order.get(r.get("quantization", ""), 999))
BENCHMARKS_KNOWN = ("gsm8k", "math500", "gpqa")
def _benchmark_from_filename(fname: str) -> str:
"""Extract the benchmark from a metrics filename.
Filenames written by stepprobe.metrics look like
{model}_{quant}_{benchmark}_run{N}_metrics.json
and both model and quant contain underscores, so we match the benchmark
by a known-values list rather than by position.
"""
import re
for bench in BENCHMARKS_KNOWN:
if re.search(rf"_{bench}_run\d+_metrics\.json$", fname):
return bench
return "unknown"
def load_all_metrics(metrics_dir: str) -> List[dict]:
"""Load all metrics JSON files from a directory, annotating each with its benchmark."""
results = []
for f in sorted(glob.glob(os.path.join(metrics_dir, "*_metrics.json"))):
data = load_json(f)
data["benchmark"] = _benchmark_from_filename(os.path.basename(f))
results.append(data)
return results
def fig1_ssr_curves(results: List[dict], output_path: str, title_suffix: str = ""):
"""
Figure 1: Step Survival Rate curves.
Base quants drawn solid, restored variants drawn dashed in the same hue.
"""
fig, ax = plt.subplots(figsize=(8.5, 5.2))
for r in _sorted_results(results):
quant = r.get("quantization", "")
ssr = r.get("ssr_curve", [])
if not ssr:
continue
color = COLORS.get(quant, "#888888")
is_restored = quant.endswith("_restored")
ax.plot(
range(len(ssr)), ssr,
label=_pretty_label(quant),
color=color,
linewidth=2.0 if not is_restored else 2.0,
linestyle="--" if is_restored else "-",
alpha=0.95,
)
ax.set_xlabel("Reasoning step depth")
ax.set_ylabel("Fraction of runs still correct")
title = "Step survival — how reasoning degrades with depth"
if title_suffix:
title = f"{title}\n{title_suffix}"
ax.set_title(title)
ax.set_ylim(0, 1.05)
ax.legend(loc="upper right", frameon=True, framealpha=0.9, ncol=1)
ax.grid(True, alpha=0.25)
ax.spines["top"].set_visible(False)
ax.spines["right"].set_visible(False)
plt.savefig(output_path)
plt.close()
print(f" Fig 1 saved: {output_path}")
def fig2_error_type_heatmap(results: List[dict], output_path: str, title_suffix: str = ""):
"""
Figure 2: Error type distribution per quantization.
Paired base/restored rows so you can read "did restoration reduce the
conceptual/logical error share?" at a glance.
"""
error_types = ["conceptual", "methodological", "executional", "logical"]
sorted_res = _sorted_results([r for r in results if r.get("error_type_dist")])
if not sorted_res:
print(" [SKIP] No error distribution data for heatmap")
return
quants = [r["quantization"] for r in sorted_res]
data = np.array([[r["error_type_dist"].get(e, 0) for e in error_types] for r in sorted_res])
fig, ax = plt.subplots(figsize=(7.5, 0.55 * len(quants) + 1.6))
im = ax.imshow(data, cmap="YlOrRd", aspect="auto", vmin=0, vmax=max(0.6, data.max()))
ax.set_xticks(range(len(error_types)))
ax.set_xticklabels([e.capitalize() for e in error_types])
ax.set_yticks(range(len(quants)))
ax.set_yticklabels([_pretty_label(q) for q in quants])
for i in range(len(quants)):
for j in range(len(error_types)):
val = data[i, j]
color = "white" if val > 0.35 else "black"
ax.text(j, i, f"{val:.0%}", ha="center", va="center", color=color, fontsize=10)
title = "Error type distribution by quantization"
if title_suffix:
title = f"{title}\n{title_suffix}"
ax.set_title(title)
plt.colorbar(im, ax=ax, label="Fraction of errors", fraction=0.04, pad=0.04)
plt.savefig(output_path)
plt.close()
print(f" Fig 2 saved: {output_path}")
def _plot_paired_bars(ax, results, value_fn, *, ylabel, percent=False):
"""Draw grouped (base, restored) bars for the three quant methods.
Each method family (awq, gptq, bnb_nf4) gets one group on the x-axis.
Within a group: two adjacent bars — base (darker) and restored (lighter).
Bars are annotated with their numeric value.
"""
families = [("awq_w4", "awq_w4_restored", "AWQ w4"),
("gptq_w4", "gptq_w4_restored", "GPTQ w4"),
("bnb_nf4_w4", "bnb_nf4_w4_restored", "BnB NF4")]
by_q = {r.get("quantization", ""): r for r in results}
x = np.arange(len(families))
width = 0.36
base_vals, restored_vals = [], []
for base_q, rest_q, _ in families:
base_vals.append(value_fn(by_q.get(base_q)))
restored_vals.append(value_fn(by_q.get(rest_q)))
base_colors = [COLORS.get(families[i][0], "#888888") for i in range(len(families))]
rest_colors = [COLORS.get(families[i][1], "#BBBBBB") for i in range(len(families))]
b1 = ax.bar(x - width / 2, base_vals, width, color=base_colors,
edgecolor="white", linewidth=0.6, label="Quantized")
b2 = ax.bar(x + width / 2, restored_vals, width, color=rest_colors,
edgecolor="white", linewidth=0.6, label="Restored")
def fmt(v):
if v is None or (isinstance(v, float) and (v != v)):
return ""
return f"{v:.0%}" if percent else f"{v:.2f}"
for bars, vals in ((b1, base_vals), (b2, restored_vals)):
for bar, v in zip(bars, vals):
if v is None:
continue
ax.text(bar.get_x() + bar.get_width() / 2,
bar.get_height() + (0.01 if percent else 0.02),
fmt(v), ha="center", va="bottom", fontsize=9)
ax.set_xticks(x)
ax.set_xticklabels([f[2] for f in families])
ax.set_ylabel(ylabel)
ax.grid(True, axis="y", alpha=0.25)
ax.spines["top"].set_visible(False)
ax.spines["right"].set_visible(False)
ax.legend(loc="best", frameon=True, framealpha=0.9)
def fig3_ffs_distribution(results: List[dict], output_path: str, title_suffix: str = ""):
"""Figure 3: First Failure Step — grouped (base, restored) bars per method."""
fig, ax = plt.subplots(figsize=(8, 5))
def get_ffs(r):
if not r:
return None
v = r.get("avg_ffs", None)
if v is None or v == float("inf"):
return None
return v
_plot_paired_bars(ax, results, get_ffs, ylabel="Avg first failure step")
title = "Where does reasoning first break? (higher = better)"
if title_suffix:
title = f"{title}\n{title_suffix}"
ax.set_title(title)
plt.savefig(output_path)
plt.close()
print(f" Fig 3 saved: {output_path}")
def fig4_accuracy_degradation(results: List[dict], fp16_acc: float, output_path: str, title_suffix: str = ""):
"""Figure 4: Accuracy — grouped (base, restored) bars per method, with FP16 line."""
fig, ax = plt.subplots(figsize=(8, 5))
def get_acc(r):
return r.get("accuracy") if r else None
_plot_paired_bars(ax, results, get_acc, ylabel="Accuracy", percent=True)
if fp16_acc is not None:
ax.axhline(y=fp16_acc, color=COLORS["fp16"], linestyle="--",
linewidth=1.5, alpha=0.8, label="FP16 reference")
# Re-draw legend so the FP16 line is included.
ax.legend(loc="best", frameon=True, framealpha=0.9)
ax.set_ylim(0, max(1.0, (ax.get_ylim()[1] or 0) + 0.05))
title = "Accuracy — quantized vs restored (higher = better)"
if title_suffix:
title = f"{title}\n{title_suffix}"
ax.set_title(title)
plt.savefig(output_path)
plt.close()
print(f" Fig 4 saved: {output_path}")
def fig5_cascade_rate(results: List[dict], output_path: str, title_suffix: str = ""):
"""Figure 5: Error Cascade Rate — grouped (base, restored) bars per method."""
fig, ax = plt.subplots(figsize=(8, 5))
def get_ecr(r):
return r.get("ecr") if r else None
_plot_paired_bars(ax, results, get_ecr, ylabel="Error cascade rate", percent=True)
ax.set_ylim(0, 1.05)
title = "Once reasoning breaks, how badly does it cascade? (lower = better)"
if title_suffix:
title = f"{title}\n{title_suffix}"
ax.set_title(title)
plt.savefig(output_path)
plt.close()
print(f" Fig 5 saved: {output_path}")
def fig6_restoration_comparison(before: dict, after: dict, output_path: str):
"""
Figure 6: Before/after restoration for a single quantization method.
Uses the paired colors from COLORS so the "before" and "after" bars match
the hue used for that method in figs 3/4/5.
"""
metrics = ["accuracy", "avg_ffs", "ecr"]
labels = ["Accuracy", "Avg FFS\n(higher = better)", "ECR\n(lower = better)"]
before_vals = [before.get(m, 0) for m in metrics]
after_vals = [after.get(m, 0) for m in metrics]
base_q = before.get("quantization", "gptq_w4")
rest_q = after.get("quantization", f"{base_q}_restored")
bar_colors = [COLORS.get(base_q, "#888888"), COLORS.get(rest_q, "#BBBBBB")]
fig, axes = plt.subplots(1, 3, figsize=(12, 4))
for i, (ax, label, bv, av) in enumerate(zip(axes, labels, before_vals, after_vals)):
vals = [bv, av]
bars = ax.bar([0, 1], vals, color=bar_colors, width=0.55, edgecolor="white")
ax.set_xticks([0, 1])
ax.set_xticklabels([_pretty_label(base_q), _pretty_label(rest_q)])
ax.set_title(label)
ax.grid(True, axis="y", alpha=0.25)
ax.spines["top"].set_visible(False)
ax.spines["right"].set_visible(False)
for bar, v in zip(bars, vals):
fmt = f"{v:.1%}" if i != 1 else f"{v:.2f}"
ax.text(bar.get_x() + bar.get_width() / 2, bar.get_height() + 0.01,
fmt, ha="center", va="bottom", fontsize=10)
plt.suptitle(f"Restoration effect: {_pretty_label(base_q)} → {_pretty_label(rest_q)}",
fontsize=13, y=1.02)
plt.tight_layout()
plt.savefig(output_path)
plt.close()
print(f" Fig 6 saved: {output_path}")
def generate_all_figures(metrics_dir: str, output_dir: str, fp16_acc: float = 0.85,
model_filter: str = None, benchmark_filter: str = None):
"""Generate figures grouped by (model, benchmark).
Each group produces one set of figures under output_dir/<model>/<benchmark>/
so that the quantization lines/bars on each plot compare apples-to-apples
(same model, same benchmark).
If model_filter/benchmark_filter are set, only matching groups are rendered.
"""
from collections import defaultdict
os.makedirs(output_dir, exist_ok=True)
results = load_all_metrics(metrics_dir)
if not results:
print("No metrics found. Run the evaluation pipeline first.")
print(f" Expected: {metrics_dir}/*_metrics.json")
return
print(f"Loaded {len(results)} metric files")
# Group by (model, benchmark). Without grouping, a single figure mixed
# together entries from every model and every benchmark and was unreadable.
groups = defaultdict(list)
for r in results:
model = r.get("model", "unknown") or "unknown"
bench = r.get("benchmark", "unknown") or "unknown"
if model_filter and model != model_filter:
continue
if benchmark_filter and bench != benchmark_filter:
continue
groups[(model, bench)].append(r)
if not groups:
print("No metric files matched the given --model/--benchmark filters.")
return
for (model, bench), group in sorted(groups.items()):
subdir = os.path.join(output_dir, model, bench)
os.makedirs(subdir, exist_ok=True)
suffix = f"{model} · {bench}"
print(f"\n[{model} / {bench}] {len(group)} quant variants")
fig1_ssr_curves(group, os.path.join(subdir, "fig1_ssr_curves.pdf"), title_suffix=suffix)
fig2_error_type_heatmap(group, os.path.join(subdir, "fig2_error_heatmap.pdf"), title_suffix=suffix)
fig3_ffs_distribution(group, os.path.join(subdir, "fig3_ffs_distribution.pdf"), title_suffix=suffix)
fig4_accuracy_degradation(group, fp16_acc, os.path.join(subdir, "fig4_accuracy_degradation.pdf"), title_suffix=suffix)
fig5_cascade_rate(group, os.path.join(subdir, "fig5_cascade_rate.pdf"), title_suffix=suffix)
# Fig 6: before/after restoration within this (model, benchmark).
for r in group:
q = r.get("quantization", "")
if not q.endswith("_restored"):
continue
base_q = q[: -len("_restored")]
before = next((x for x in group if x.get("quantization") == base_q), None)
if before:
out = os.path.join(subdir, f"fig6_restoration_{base_q}.pdf")
fig6_restoration_comparison(before, r, out)
print(f"\nAll figures saved under {output_dir}/<model>/<benchmark>/")
# ============================================================
# Demo with synthetic data (for testing)
# ============================================================
def generate_demo_figures(output_dir: str):
"""Generate demo figures with synthetic data for testing the visualization."""
os.makedirs(output_dir, exist_ok=True)
# Synthetic results
demo_results = [
{
"model": "DeepSeek-R1-Distill-Qwen-7B",
"quantization": "awq_w4",
"accuracy": 0.72,
"accuracy_delta": -0.13,
"avg_ffs": 3.2,
"median_ffs": 3.0,
"ffs_std": 1.8,
"ecr": 0.78,
"ssr_curve": [0.95, 0.88, 0.79, 0.68, 0.55, 0.45, 0.38, 0.32, 0.28, 0.25,
0.23, 0.21, 0.20, 0.19, 0.18, 0.17, 0.16, 0.15, 0.14, 0.13],
"error_type_dist": {"conceptual": 0.12, "methodological": 0.25, "executional": 0.48, "logical": 0.15},
},
{
"model": "DeepSeek-R1-Distill-Qwen-7B",
"quantization": "awq_w3",
"accuracy": 0.58,
"accuracy_delta": -0.27,
"avg_ffs": 2.1,
"median_ffs": 2.0,
"ffs_std": 1.3,
"ecr": 0.89,
"ssr_curve": [0.90, 0.75, 0.58, 0.42, 0.30, 0.22, 0.17, 0.14, 0.12, 0.10,
0.09, 0.08, 0.07, 0.06, 0.05, 0.05, 0.04, 0.04, 0.03, 0.03],
"error_type_dist": {"conceptual": 0.28, "methodological": 0.22, "executional": 0.35, "logical": 0.15},
},
{
"model": "DeepSeek-R1-Distill-Qwen-7B",
"quantization": "gptq_w4",
"accuracy": 0.74,
"accuracy_delta": -0.11,
"avg_ffs": 3.5,
"median_ffs": 3.0,
"ffs_std": 2.0,
"ecr": 0.75,
"ssr_curve": [0.96, 0.90, 0.82, 0.72, 0.60, 0.50, 0.42, 0.36, 0.31, 0.27,
0.24, 0.22, 0.20, 0.19, 0.18, 0.17, 0.16, 0.15, 0.14, 0.13],
"error_type_dist": {"conceptual": 0.10, "methodological": 0.20, "executional": 0.55, "logical": 0.15},
},
{
"model": "DeepSeek-R1-Distill-Qwen-7B",
"quantization": "bnb_nf4",
"accuracy": 0.70,
"accuracy_delta": -0.15,
"avg_ffs": 3.0,
"median_ffs": 3.0,
"ffs_std": 1.9,
"ecr": 0.80,
"ssr_curve": [0.94, 0.86, 0.76, 0.64, 0.52, 0.42, 0.35, 0.30, 0.26, 0.23,
0.21, 0.19, 0.18, 0.17, 0.16, 0.15, 0.14, 0.13, 0.12, 0.11],
"error_type_dist": {"conceptual": 0.15, "methodological": 0.23, "executional": 0.45, "logical": 0.17},
},
]
fig1_ssr_curves(demo_results, os.path.join(output_dir, "fig1_ssr_curves.png"))
fig2_error_type_heatmap(demo_results, os.path.join(output_dir, "fig2_error_heatmap.png"))
fig3_ffs_distribution(demo_results, os.path.join(output_dir, "fig3_ffs_distribution.png"))
fig4_accuracy_degradation(demo_results, 0.85, os.path.join(output_dir, "fig4_accuracy_degradation.png"))
fig5_cascade_rate(demo_results, os.path.join(output_dir, "fig5_cascade_rate.png"))
# Demo restoration
before = demo_results[2] # gptq_w4
after = {"accuracy": 0.82, "avg_ffs": 5.1, "ecr": 0.45}
fig6_restoration_comparison(before, after, os.path.join(output_dir, "fig6_restoration.png"))
print(f"\nDemo figures saved to {output_dir}")
if __name__ == "__main__":
import argparse
parser = argparse.ArgumentParser(description="Generate paper figures")
parser.add_argument("--metrics", default=None, help="Metrics directory")
parser.add_argument("--output", default="figures/", help="Output directory")
parser.add_argument("--fp16-acc", type=float, default=0.85)
parser.add_argument("--model", default=None, help="Only render figures for this model tag")
parser.add_argument("--benchmark", default=None, help="Only render figures for this benchmark")
parser.add_argument("--demo", action="store_true", help="Generate demo figures with synthetic data")
args = parser.parse_args()
if args.demo:
generate_demo_figures(args.output)
elif args.metrics:
generate_all_figures(args.metrics, args.output, fp16_acc=args.fp16_acc,
model_filter=args.model, benchmark_filter=args.benchmark)
else:
print("Specify --metrics <dir> or --demo")
|