File size: 8,618 Bytes
05ab4b5 | 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 | """Generate all data figures for the paper (PDF + PNG preview).
Fig 3 (audit): (a) 2x4 small-multiple FQ x J-Access scatter, each method
highlighted against the full grid in gray;
(b) layer profile with the workspace band shaded.
Fig 4 (predict): 4x2 small multiples, one per unlearning method: nine
checkpoints of pre-attack J-Access vs excess revival
with a Theil-Sen fit.
Fig 5 (goodhart): two panels sharing lambda: J-Access_K + UDS (left), revival
(right); three-seed error bars.
Palette: dataviz reference palette (light), <=3 categorical hues per
all-pairs panel; identity via panel titles / direct labels, never color
alone. One axis per panel.
"""
import json
import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt
import numpy as np
from matplotlib.ticker import MaxNLocator
from scipy import stats
RES = "/workspace/jspace-unlearning/results"
OUT = "/workspace/jspace-unlearning/paper/figs"
BLUE, ORANGE, AQUA = "#2a78d6", "#eb6834", "#1baf7a"
INK, MUTED, GRIDC = "#0b0b0b", "#52514e", "#e6e5e1"
GRAY_PT = "#c9c8c2"
plt.rcParams.update({
"font.size": 7.5, "axes.titlesize": 8, "axes.labelsize": 7.5,
"xtick.labelsize": 7, "ytick.labelsize": 7, "legend.fontsize": 7,
"axes.edgecolor": MUTED, "axes.linewidth": 0.6,
"xtick.color": MUTED, "ytick.color": MUTED,
"axes.labelcolor": INK, "text.color": INK,
"figure.dpi": 200, "savefig.bbox": "tight",
})
def style(ax):
for s in ("top", "right"):
ax.spines[s].set_visible(False)
def save(fig, name):
fig.savefig(f"{OUT}/{name}.pdf")
fig.savefig(f"{OUT}/{name}.png")
plt.close(fig)
print("saved", name)
# ------------------------------------------------------------------ fig 3
def fig3():
grid = json.load(open(f"{RES}/p0a_grid398.json"))
# drop runs whose KS p-value underflowed to 0 (would pile up at the
# clip value); the paper notes their omission
models = {k: v for k, v in grid["models"].items() if v["fq"] > 1e-30}
jf, jg = grid["digest"]["j_full"], grid["digest"]["j_gold"]
lfq = {k: np.log10(v["fq"]) for k, v in models.items()}
methods = sorted({v["method"] for v in models.values()})
fig, axes = plt.subplots(2, 4, figsize=(7.0, 3.6), sharex=True,
sharey=True, constrained_layout=True)
for ax, meth in zip(axes.flat, methods):
style(ax)
xs_all = [lfq[k] for k in models]
ys_all = [models[k]["jocc10"] for k in models]
ax.scatter(xs_all, ys_all, s=4, color=GRAY_PT, linewidths=0,
zorder=2, rasterized=True)
sel = {k: v for k, v in models.items() if v["method"] == meth}
succ = {k: v for k, v in sel.items()
if v["fq"] > 0.05 and v["mu"] >= 0.9 * 0.5994}
ax.scatter([lfq[k] for k in sel], [v["jocc10"] for v in sel.values()],
s=7, color=BLUE, linewidths=0, zorder=3)
ax.scatter([lfq[k] for k in succ], [v["jocc10"] for v in succ.values()],
s=16, facecolors="none", edgecolors=INK, linewidths=0.7,
zorder=4)
ax.axhline(jf, color=MUTED, lw=0.7, ls=(0, (4, 2)), zorder=1)
ax.axhline(jg, color=MUTED, lw=0.7, ls=(0, (1, 1.5)), zorder=1)
ax.set_title(meth, pad=2)
ax.set_xlim(-31, 1.5)
ax.set_ylim(-0.03, 0.85)
fig.supxlabel(r"$\log_{10}$ Forget Quality $\;\longrightarrow\;$"
" better behavioral forgetting", fontsize=7.5)
fig.supylabel(r"Raw Target Access Rate $A_{10}$", fontsize=7.5)
axes[0, 0].text(-30, jf + 0.02, r"$\theta_0$", fontsize=7, color=MUTED)
axes[0, 0].text(-30, jg + 0.02, r"$\theta_g$", fontsize=7, color=MUTED)
axes[0, 0].annotate("lower residual\naccess", xy=(-23, 0.01),
xytext=(-23, 0.13), fontsize=6, color=MUTED,
ha="center", va="bottom",
arrowprops=dict(arrowstyle="->", color=MUTED, lw=0.7))
handles = [
plt.Line2D([], [], marker="o", ls="", ms=3, color=GRAY_PT,
label="all runs (pooled)"),
plt.Line2D([], [], marker="o", ls="", ms=3.5, color=BLUE,
label="this method's runs"),
plt.Line2D([], [], marker="o", ls="", ms=5, mfc="none", mec=INK,
mew=0.7, label="passes FQ + utility gates"),
plt.Line2D([], [], color=MUTED, lw=0.9, ls=(0, (4, 2)),
label=r"Original model $\theta_0$"),
plt.Line2D([], [], color=MUTED, lw=0.9, ls=(0, (1, 1.5)),
label=r"Retain-only model $\theta_g$"),
]
fig.legend(handles=handles, ncol=5, frameon=False, fontsize=6.5,
loc="outside upper center", columnspacing=1.0,
handletextpad=0.4, handlelength=1.6)
save(fig, "fig3a_grid_scatter")
prof = json.load(open(f"{RES}/p2b_layer_profile.json"))
fig, ax = plt.subplots(figsize=(3.3, 2.1), constrained_layout=True)
style(ax)
ax.axvspan(11, 14, color=GRIDC, alpha=0.6, zorder=0)
series = [("full", "full", BLUE, "-"),
("retain90", "gold (retain-only)", MUTED, (0, (4, 2))),
("NPO", "NPO (best FQ)", ORANGE, "-"),
("RMU", "RMU", AQUA, "-")]
for key, label, color, ls in series:
p = prof[key]["jocc10_by_layer_prefix"]
xs = [int(k) for k in p if k != "15"]
ys = [p[str(x)] for x in xs]
ax.plot(xs, ys, color=color, lw=1.4, ls=ls, zorder=3)
ax.annotate(label, (xs[-1], ys[-1]), textcoords="offset points",
xytext=(3, 0), fontsize=6.5, color=INK, va="center")
ax.set_xlabel("lens layer")
ax.set_ylabel("J-Access (per layer)")
ax.set_xlim(0, 17)
ax.set_xticks(range(0, 15, 2))
ax.text(12.5, 0.55, "workspace\nband", fontsize=6.5, color=MUTED,
ha="center")
save(fig, "fig3b_layer_profile")
# ------------------------------------------------------------------ fig 4
def fig4():
h2 = json.load(open(f"{RES}/h2_ckpt_stats.json"))
methods = sorted({k.split("_")[0] for k in h2})
fig, axes = plt.subplots(4, 2, figsize=(3.1, 4.4), sharey=True,
constrained_layout=True)
for ax, meth in zip(axes.flat, methods):
style(ax)
xs, ye = map(np.array, zip(*[
(v["jocc_v2_mean"], v["excess_rate"]) for k, v in h2.items()
if k.split("_")[0] == meth]))
ax.axhline(0, color=GRAY_PT, lw=0.6)
ax.scatter(xs, ye, s=8, color=BLUE, linewidths=0, zorder=3)
slope, intercept, _, _ = stats.theilslopes(ye, xs)
xf = np.array([xs.min(), xs.max()])
ax.plot(xf, intercept + slope * xf, color=INK, lw=0.9,
ls=(0, (4, 2)), zorder=2)
ax.set_title(meth, pad=2)
ax.xaxis.set_major_locator(MaxNLocator(3))
ax.set_ylim(-0.6, 0.5)
ax.set_yticks([-0.4, 0, 0.4])
fig.supxlabel("pre-attack J-Access", fontsize=7.5)
fig.supylabel("excess revival", fontsize=7.5)
save(fig, "fig4_prediction")
# ------------------------------------------------------------------ fig 5
def fig5():
seeds = json.load(open(f"{RES}/p1b_seeds.json"))
uds = json.load(open(f"{RES}/h3_uds_comparison.json"))
lams = [0, 5, 10]
def agg(field):
m, s = [], []
for lam in lams:
v = [r[field] for r in seeds.values() if r["lam"] == lam]
m.append(np.mean(v))
s.append(np.std(v, ddof=1))
return np.array(m), np.array(s)
jm, js = agg("JOcc2")
rm, rs = agg("within_revive")
um = [uds[f"WDd_l{lam}"]["UDS"] for lam in lams]
fig, (a, b) = plt.subplots(1, 2, figsize=(3.92, 2.4), sharex=True,
constrained_layout=True)
# all text (labels, tick labels, annotations) left off: added by hand later
for ax in (a, b):
style(ax)
ax.set_xticks(lams)
ax.set_xticklabels([])
a.errorbar(lams, jm, yerr=js, color=BLUE, lw=1.4, marker="o", ms=3.5,
capsize=2, elinewidth=0.8)
a.plot(lams, um, color=MUTED, lw=1.2, ls=(0, (4, 2)), marker="s", ms=3)
a.set_ylim(0.45, 0.85)
a.set_yticks([0.5, 0.6, 0.7, 0.8])
a.set_yticklabels([])
b.errorbar(lams, rm, yerr=rs, color=ORANGE, lw=1.4, marker="o", ms=3.5,
capsize=2, elinewidth=0.8)
b.set_ylim(0.24, 0.42)
b.set_yticks([0.25, 0.30, 0.35, 0.40])
b.set_yticklabels([])
save(fig, "fig5_goodhart")
if __name__ == "__main__":
import os
os.makedirs(OUT, exist_ok=True)
fig3()
fig4()
fig5()
|