File size: 5,991 Bytes
d428e08 | 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 | """Render the Llama LR sweep figure (paper fig 8).
ΔAccuracy of the restored adapter vs the quantized baseline, per learning
rate. Zero line is the "no effect" reference — below zero = restoration
made it worse. The figure is laid out so the story ("lower LR recovers
more of the gap") reads in one pass.
"""
import argparse
import json
import os
import re
import sys
import matplotlib as mpl
import matplotlib.pyplot as plt
import numpy as np
sys.path.insert(0, os.path.dirname(os.path.abspath(__file__)))
mpl.rcParams.update({
"font.family": "sans-serif",
"font.sans-serif": ["Inter", "Helvetica Neue", "Arial", "DejaVu Sans"],
"font.size": 9,
"axes.labelsize": 10,
"xtick.labelsize": 8.5,
"ytick.labelsize": 8.5,
"legend.fontsize": 8,
"legend.frameon": False,
"figure.dpi": 200,
"savefig.dpi": 400,
"savefig.bbox": "tight",
"pdf.fonttype": 42,
"ps.fonttype": 42,
"axes.linewidth": 0.7,
"axes.spines.top": False,
"axes.spines.right": False,
})
LR_PARSE = re.compile(r"^lr(.+)$")
METHOD_COLOR = {"awq_w4": "#4E79A7", "gptq_w4": "#E15759", "bnb_nf4_w4": "#59A14F"}
GAIN = "#2E7D32"
LOSS = "#C03A2B"
def _load_outcomes(jsonl):
if not os.path.exists(jsonl):
return None
out = {}
with open(jsonl) as f:
for line in f:
t = json.loads(line)
out[t["problem_id"]] = 1.0 if t.get("is_correct_final") else 0.0
return out
def paired_delta_ci(base_out, rest_out, n_boot=5000):
ids = sorted(set(base_out) & set(rest_out))
if not ids:
return None
b = np.array([base_out[i] for i in ids])
r = np.array([rest_out[i] for i in ids])
n = len(ids)
rng = np.random.default_rng(0)
deltas = np.empty(n_boot)
for i in range(n_boot):
idx = rng.integers(0, n, size=n)
deltas[i] = r[idx].mean() - b[idx].mean()
obs = float(r.mean() - b.mean())
return obs * 100, float(np.percentile(deltas, 2.5)) * 100, float(np.percentile(deltas, 97.5)) * 100
def _tag_to_lr(tag: str) -> float:
# lr5e_5 -> 5e-5 etc. The sweep script encodes "." as "_" and "-" as "_".
s = tag.replace("__", "-").replace("_", "-")
try:
return float(s)
except ValueError:
return float("nan")
def main():
parser = argparse.ArgumentParser()
parser.add_argument("--sweep-root", required=True)
parser.add_argument("--model", required=True)
parser.add_argument("--quant", required=True)
parser.add_argument("--benchmark", required=True)
parser.add_argument("--metrics", required=True)
parser.add_argument("--output", required=True)
args = parser.parse_args()
base_diag = os.path.join("results", "diagnosis", args.quant, args.model,
f"{args.benchmark}_run0.jsonl")
base_out = _load_outcomes(base_diag)
if base_out is None:
print(f" [ERROR] base diagnosed jsonl missing: {base_diag}")
return
rows = []
for entry in sorted(os.listdir(args.sweep_root)):
m = LR_PARSE.match(entry)
if not m:
continue
diag = os.path.join(args.sweep_root, entry, "diagnosis",
f"{args.benchmark}_run0.jsonl")
rest_out = _load_outcomes(diag)
if rest_out is None:
continue
ci = paired_delta_ci(base_out, rest_out)
if ci is None:
continue
lr_val = _tag_to_lr(m.group(1))
rows.append((lr_val, ci))
if not rows:
print("No LR sweep data found.")
return
rows.sort(key=lambda r: r[0])
lrs = [r[0] for r in rows]
deltas = [r[1][0] for r in rows]
los = [r[1][1] for r in rows]
his = [r[1][2] for r in rows]
fig, ax = plt.subplots(figsize=(5.3, 3.2), constrained_layout=True)
# Shade the "below-zero = regression" region in a faint red.
ys_needed = deltas + los + his + [0]
ymin, ymax = min(ys_needed) - 3, max(ys_needed) + 4
ax.axhspan(ymin, 0, color=LOSS, alpha=0.06, zorder=0)
ax.axhspan(0, ymax, color=GAIN, alpha=0.06, zorder=0)
# Zero reference line.
ax.axhline(0, color="#555555", linewidth=0.8, zorder=1)
x = np.arange(len(lrs))
for i, (d, lo, hi) in enumerate(zip(deltas, los, his)):
color = GAIN if d >= 0 else LOSS
ax.plot([x[i], x[i]], [lo, hi], color=color, linewidth=1.4, alpha=0.6,
zorder=2, solid_capstyle="butt")
ax.plot(x[i], d, "o", markersize=8, color=color, markeredgecolor="white",
markeredgewidth=1.2, zorder=3)
ax.annotate(f"{d:+.1f} pp", xy=(x[i], d), xytext=(0, 12 if d >= 0 else -16),
textcoords="offset points", ha="center",
va="bottom" if d >= 0 else "top",
fontsize=9.5, color=color, fontweight="bold")
# Edge-of-plot labels for the shaded regions.
ax.text(-0.45, ymax * 0.92, "restoration helps",
ha="left", va="top", fontsize=8, color=GAIN,
style="italic")
ax.text(-0.45, ymin + 0.5, "restoration hurts",
ha="left", va="bottom", fontsize=8, color=LOSS,
style="italic")
ax.set_xticks(x)
ax.set_xticklabels([f"{lr:g}" for lr in lrs])
ax.set_xlabel("QLoRA learning rate")
ax.set_ylabel(r"$\Delta$Accuracy vs. quantized baseline (pp)")
ax.yaxis.grid(True, linewidth=0.4, color="#DDDDDD")
ax.set_axisbelow(True)
ax.set_ylim(ymin, ymax)
ax.set_xlim(-0.55, len(lrs) - 0.45)
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=8, color="#555")
os.makedirs(os.path.dirname(args.output), exist_ok=True)
fig.savefig(args.output)
plt.close(fig)
print(f" Paper fig 8 (Llama LR sweep) saved: {args.output}")
if __name__ == "__main__":
main()
|