StepProbe / scripts /make_ablation_figure.py
Akiyue's picture
Add files using upload-large-folder tool
3ccaf5a verified
Raw
History Blame Contribute Delete
11.9 kB
"""Render the dataset-size ablation figure (paper fig 6).
Reads restoration results produced by run_ablation.sh and plots accuracy of
the restored model as the silver-bullet dataset size N varies, with two
reference lines:
- quantized baseline (no restoration) — from results/metrics/
- FP16 upper bound (if available) — from results/segmented/fp16/
Style matches the other paper figures (Q1 conventions — Type-42 fonts, muted
Tableau palette, no inline title).
"""
import argparse
import json
import os
import re
from typing import List, Optional
import matplotlib as mpl
import matplotlib.pyplot as plt
import numpy as np
# ---------------------------------------------------------------------------
# Style (matches make_paper_figures.py exactly)
# ---------------------------------------------------------------------------
mpl.rcParams.update({
"font.family": "sans-serif",
"font.sans-serif": ["Inter", "Helvetica Neue", "Arial", "DejaVu Sans"],
"font.size": 8,
"axes.titlesize": 9,
"axes.labelsize": 9,
"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.spines.top": False,
"axes.spines.right": False,
"xtick.major.width": 0.6,
"ytick.major.width": 0.6,
"grid.color": "#EAEAEA",
"grid.linewidth": 0.5,
})
METHOD_COLOR = {
"awq_w4": "#4E79A7",
"gptq_w4": "#E15759",
"bnb_nf4_w4": "#59A14F",
}
FP16_COLOR = "#333333"
BASE_COLOR = "#888888"
GREY_TEXT = "#555555"
# ---------------------------------------------------------------------------
# Accuracy helpers
# ---------------------------------------------------------------------------
def accuracy_from_diagnosed(jsonl_path: str) -> Optional[float]:
if not os.path.exists(jsonl_path):
return None
n, c = 0, 0
with open(jsonl_path) as f:
for line in f:
t = json.loads(line)
n += 1
if t.get("is_correct_final"):
c += 1
return c / n if n else None
def ci_from_diagnosed(jsonl_path: str, n_boot: int = 2000) -> Optional[tuple]:
"""Return (acc, ci_lo, ci_hi) from per-problem bootstrap."""
if not os.path.exists(jsonl_path):
return None
vec = []
with open(jsonl_path) as f:
for line in f:
t = json.loads(line)
vec.append(1.0 if t.get("is_correct_final") else 0.0)
if not vec:
return None
vec = np.array(vec)
rng = np.random.default_rng(0)
samples = np.empty(n_boot)
for i in range(n_boot):
idx = rng.integers(0, len(vec), size=len(vec))
samples[i] = vec[idx].mean()
lo, hi = np.percentile(samples, [2.5, 97.5])
return float(vec.mean()), float(lo), float(hi)
def load_baseline_acc(metrics_dir, model, quant, bench):
path = os.path.join(metrics_dir, f"{model}_{quant}_{bench}_run0_metrics.json")
if not os.path.exists(path):
return None
with open(path) as f:
return json.load(f).get("accuracy")
def load_fp16_acc_from_segmented(segmented_dir, model, bench) -> Optional[float]:
"""FP16 accuracy from segmented jsonl.
Segment doesn't write `is_correct_final`, so we compute accuracy ourselves
by comparing `final_answer` to the benchmark's gold answer. Returns None
if anything is missing — caller treats None as "skip the FP16 reference
line" rather than silently plotting 0%."""
path = os.path.join(segmented_dir, "fp16", model, f"{bench}_run0.jsonl")
if not os.path.exists(path):
return None
# Load gold answers for the benchmark
try:
from datasets import load_dataset
if bench == "gsm8k":
ds = load_dataset("openai/gsm8k", "main", split="test")
golds = {f"gsm8k_{i}": ex["answer"].split("####")[-1].strip()
for i, ex in enumerate(ds)}
elif bench == "math500":
ds = load_dataset("HuggingFaceH4/MATH-500", split="test")
golds = {f"math500_{i}": ex["answer"] for i, ex in enumerate(ds)}
elif bench == "gpqa":
ds = load_dataset("Idavidrein/gpqa", "gpqa_diamond", split="train")
golds = {f"gpqa_{i}": ex.get("Correct Answer", "") for i, ex in enumerate(ds)}
else:
return None
except Exception:
return None
def norm(s: str) -> str:
s = (s or "").strip().strip("$").replace(" ", "").replace(",", "")
return s.lower()
n, c = 0, 0
with open(path) as f:
for line in f:
t = json.loads(line)
pid = t.get("problem_id")
gold = golds.get(pid, "")
pred = t.get("final_answer", "")
n += 1
if gold and pred and norm(gold) == norm(pred):
c += 1
return c / n if n else None
# ---------------------------------------------------------------------------
# Main
# ---------------------------------------------------------------------------
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--ablation-root", required=True,
help="results/ablation/<model>_<quant>")
parser.add_argument("--model", required=True)
parser.add_argument("--quant", required=True)
parser.add_argument("--benchmark", required=True)
parser.add_argument("--metrics", required=True,
help="results/metrics — used for the quantized baseline")
parser.add_argument("--segmented", default="results/segmented",
help="results/segmented — used for FP16 upper bound (optional)")
parser.add_argument("--output", required=True)
args = parser.parse_args()
# Discover ablation runs on disk: n50, n100, n250, n500, ...
Ns = []
accs, ci_los, ci_his = [], [], []
for entry in sorted(os.listdir(args.ablation_root)):
m = re.match(r"n(\d+)$", entry)
if not m:
continue
N = int(m.group(1))
diag = os.path.join(args.ablation_root, entry, "diagnosis",
f"{args.benchmark}_run0.jsonl")
ci = ci_from_diagnosed(diag)
if ci is None:
print(f" [SKIP] missing diagnosed jsonl for N={N}: {diag}")
continue
Ns.append(N)
accs.append(ci[0])
ci_los.append(ci[1])
ci_his.append(ci[2])
if not Ns:
print(f"No ablation data found in {args.ablation_root}")
return
sort_idx = np.argsort(Ns)
Ns = [Ns[i] for i in sort_idx]
accs = [accs[i] for i in sort_idx]
ci_los = [ci_los[i] for i in sort_idx]
ci_his = [ci_his[i] for i in sort_idx]
base_acc = load_baseline_acc(args.metrics, args.model, args.quant, args.benchmark)
# FP16 upper-bound via LaTeX-aware eval (math_verify + SymPy). See
# scripts/eval_accuracy.py for the brace-balanced extractor that fixes
# the old naive string-match approach.
try:
from scripts.eval_accuracy import accuracy as _lv_accuracy
except ImportError:
import sys as _sys
_sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
from eval_accuracy import accuracy as _lv_accuracy
fp16_jsonl = os.path.join(args.segmented, "fp16", args.model,
f"{args.benchmark}_run0.jsonl")
fp16_acc = _lv_accuracy(fp16_jsonl, args.benchmark)
# ---------- Figure ----------
fig, ax = plt.subplots(figsize=(4.5, 3.0), constrained_layout=True)
method_color = METHOD_COLOR.get(args.quant, "#4E79A7")
# Main curve: restored accuracy vs N (with CI whiskers).
acc_pp = [a * 100 for a in accs]
lo_pp = [a * 100 for a in ci_los]
hi_pp = [a * 100 for a in ci_his]
# Per-point 95% CI whiskers (vertical error bars). Using whiskers rather
# than a fill_between band, because with only 4 discrete N values a band
# looks watery — whiskers feel more decisive.
for x, y, lo, hi in zip(Ns, acc_pp, lo_pp, hi_pp):
ax.plot([x, x], [lo, hi], color=method_color, linewidth=1.1,
alpha=0.55, zorder=2, solid_capstyle="butt")
ax.plot(Ns, acc_pp, "-", color=method_color, linewidth=1.6,
zorder=3, label="Restored")
ax.plot(Ns, acc_pp, "o", color=method_color, markersize=5.5,
markeredgecolor="white", markeredgewidth=1.0, zorder=4)
# Quantized baseline reference line
if base_acc is not None:
ax.axhline(base_acc * 100, color=BASE_COLOR, linestyle=(0, (5, 3)),
linewidth=0.9, alpha=0.9, label="Quantized (no rest.)")
# FP16 reference — only draw if we actually computed it (>0).
if fp16_acc is not None and fp16_acc > 0:
ax.axhline(fp16_acc * 100, color=FP16_COLOR, linestyle=(0, (1, 2)),
linewidth=0.9, alpha=0.9, label="FP16")
# X axis — log scale so small N values are well-spaced.
ax.set_xscale("log")
ax.set_xticks(Ns)
ax.set_xticklabels([str(n) for n in Ns])
ax.get_xaxis().set_minor_locator(mpl.ticker.NullLocator())
ax.set_xlabel("Silver-bullet dataset size (samples)")
ax.set_ylabel("Accuracy (%)")
ax.yaxis.grid(True)
ax.set_axisbelow(True)
# Y axis — zoom to the data range with a bit of padding. NOT anchored to 0,
# because clustering all data in the top fifth of the panel is a classic
# amateur Q1 tell.
ys = lo_pp + hi_pp + acc_pp
if base_acc is not None:
ys.append(base_acc * 100)
if fp16_acc is not None and fp16_acc > 0:
ys.append(fp16_acc * 100)
ymin = min(ys) - 3.0
ymax = max(ys) + 3.5
ax.set_ylim(ymin, ymax)
# Numeric labels BELOW each point so they don't collide with the corner tag.
for x, y in zip(Ns, acc_pp):
ax.annotate(f"{y:.1f}", xy=(x, y), xytext=(0, -12),
textcoords="offset points",
ha="center", va="top", fontsize=7, color=method_color)
# Corner metadata tag placed OUTSIDE the axes so it never overlaps data.
pretty_quant = {"awq_w4": "AWQ w4", "gptq_w4": "GPTQ w4", "bnb_nf4_w4": "BnB NF4"}.get(args.quant, args.quant)
ax.text(1.0, 1.02,
f"{args.model} · {pretty_quant} · {args.benchmark}",
transform=ax.transAxes, ha="right", va="bottom",
fontsize=7.5, color=GREY_TEXT)
# "+X pp over baseline" callout — anchored to the first point (N=50),
# offset UP and LEFT so the arrow doesn't cross the rising data line.
# (The previous placement at the best point shot its arrow back across
# the data in the middle of the figure.)
if base_acc is not None and len(acc_pp) > 0:
gain_at_min = acc_pp[0] - base_acc * 100
ax.annotate(
f"+{gain_at_min:.1f} pp over\nquantized baseline\nalready at N={Ns[0]}",
xy=(Ns[0], acc_pp[0]),
xytext=(0.28, 0.78), textcoords="axes fraction",
ha="left", va="top", fontsize=7, color=method_color,
arrowprops=dict(arrowstyle="->", color=method_color,
linewidth=0.7, alpha=0.7,
shrinkA=2, shrinkB=4,
connectionstyle="arc3,rad=-0.25"))
# Place legend OUTSIDE the plot (right side), so the reference lines
# (FP16 dotted, Quantized dashed) don't run through the legend labels.
ax.legend(loc="center left", bbox_to_anchor=(1.02, 0.5),
handlelength=2.0, handletextpad=0.5,
borderpad=0.3, labelspacing=0.5)
os.makedirs(os.path.dirname(args.output), exist_ok=True)
fig.savefig(args.output)
plt.close(fig)
print(f" Paper fig 6 (ablation) saved: {args.output}")
if __name__ == "__main__":
main()