Buckets:

SabaPivot
/

repro-strong-model-artifacts

Files

xet

SabaPivot/repro-strong-model-artifacts / repro-bundle /scripts /build_poster.py

SabaPivot

24 days ago

download

raw

14.1 kB

	#!/usr/bin/env python3
	"""Build the locked 60x36 reproduction poster from Posterly's 4-column template."""

	from pathlib import Path


	ROOT = Path(__file__).resolve().parents[1]
	TEMPLATE = ROOT / "vendor/posterly/templates/landscape_4col_neutral.html"
	OUT = ROOT / "poster/poster.html"

	CLAIM_1 = "claim-1-theorem-1-provides-a-teacher-to-student-risk-decomposition-into-propagated-teacher-error-student-optimization-er"
	CLAIM_2 = "claim-2-theorem-3-shows-that-under-model-mis-specification-the-distillation-efficiency-ratio-der-diverges-at-rate-omega"
	CLAIM_3 = "claim-3-theorem-4-establishes-a-weak-to-strong-generalization-guarantee-a-strong-student-model-provably-surpasses-a-weak"
	CLAIM_4 = "claim-4-theorem-5-quantifies-weak-to-strong-performance-recovery-via-performance-gap-recovered-pgr-1-o-tilde-k-dagger-2"
	CLAIM_5 = "claim-5-figures-2-and-3-empirically-validate-the-risk-inheritance-and-early-stopping-predictions-using-both-synthetic-da"

	LOCK = """<!--
	DESIGN DIRECTION — LOCKED
	Concept: Spectral field report — a calm indigo technical sheet with warm audit accents.
	Hero: one full-width outcome banner carrying the four decisive reproduction numbers.
	Canvas: 60in x 36in landscape. Density: normal. No organization logo.
	Template: posterly/templates/landscape_4col_neutral.html.
	-->"""

	CUSTOM_CSS = """
	/* Reproduction-specific, token-only additions. */
	.poster { background-image: none; }
	.header { border-bottom-width: calc(2 * var(--u)); }
	.title { max-width: calc(1500 * var(--u)); }
	.subtitle { max-width: calc(1600 * var(--u)); }
	.framework-banner .banner-stats { grid-template-columns: repeat(4, 1fr); }
	.scope-tag {
	display: inline-block; font-family: var(--font-sans); font-size: var(--fs-2);
	font-weight: 700; color: var(--emph-ink); background: var(--emph-soft);
	border: calc(1 * var(--u)) solid var(--emph); padding: calc(1 * var(--u)) calc(4 * var(--u));
	border-radius: calc(3 * var(--u) * var(--rs)); margin-bottom: calc(2 * var(--u));
	}
	.figure.compact img { max-height: calc(170 * var(--u)); object-fit: contain; }
	.figure.tall img { max-height: calc(230 * var(--u)); object-fit: contain; }
	.figure.tight img { max-height: calc(157 * var(--u)); object-fit: contain; }
	.figure.claim1-large img { height: calc(318 * var(--u)); object-fit: contain; }
	.figure.early-large img { height: calc(329 * var(--u)); object-fit: contain; }
	.figure.arch-large img { height: calc(340 * var(--u)); object-fit: contain; }
	.callout:not(.emph) strong { color: var(--accent-ink); }
	.audit-note {
	font-family: var(--font-sans); font-size: var(--fs-2); color: var(--text-secondary);
	border-top: calc(1 * var(--u)) solid var(--border-soft); padding-top: calc(3 * var(--u));
	}
	.verdict {
	font-family: var(--font-sans); font-size: var(--fs-3); font-weight: 800;
	color: var(--accent-deep); text-transform: uppercase; letter-spacing: 1px;
	}
	"""

	BODY = f"""<body>
	<div class="poster" data-measure-role="poster" data-posterly-contract="identity-v1" data-ps-identity="on">
	<svg class="ps-sprite" data-color-exempt="logo" aria-hidden="true"><defs>
	<symbol id="psReg" viewBox="0 0 100 100">
	<circle cx="50" cy="50" r="25" fill="none" stroke="currentColor" stroke-width="8"/>
	<line x1="50" y1="7" x2="50" y2="93" stroke="currentColor" stroke-width="8"/>
	<line x1="7" y1="50" x2="93" y2="50" stroke="currentColor" stroke-width="8"/>
	</symbol>
	</defs></svg>

	<header class="header" data-measure-role="header">
	<div class="venue-badge">
	<div class="vb-venue">ICML</div><div class="vb-year">2026</div><div class="vb-tag">REPRO</div>
	</div>
	<div class="title-block">
	<h1 class="title">What Makes a <span class="accent">Strong Model?</span></h1>
	<div class="subtitle">A unified spectral analysis of knowledge transfer over high-dimensional linear regression</div>
	<div class="authors-line">
	<span class="author">Wendao Wu</span> · <span class="author">Fangqing Zhang</span> ·
	<span class="author">Haihan Zhang</span> · <span class="author">Cong Fang</span>
	<span class="aff">Independent reproduction · paper #23301 · arXiv:2606.01292</span>
	</div>
	</div>
	<div class="right-block">
	<div class="qr-block">
	<img data-color-exempt="logo" src="assets/logbook-qr.png" alt="QR code to the published reproduction logbook">
	<div class="qr-label">Open evidence logbook</div>
	</div>
	</div>
	</header>

	<section class="framework-banner" data-measure-role="banner">
	<div class="fb-text"><span class="fb-label">Outcome</span>
	<strong>The spectral mechanism reproduces.</strong> Learned modes inherit teacher error; expanded horizons improve sample efficiency; selected checkpoints denoise a stronger student. Full real-world task training remains outside this scaled audit.
	</div>
	<div class="banner-stats">
	<div class="bs-item"><div class="bs-num">≤0.00996</div><div class="bs-label">tail inheritance<br>at δ² = 0.01</div></div>
	<div class="bs-item"><div class="bs-num">κ = 0.1667</div><div class="bs-label">DER slope<br>theory = fitted</div></div>
	<div class="bs-item"><div class="bs-num">5 / 5</div><div class="bs-label">signal widths<br>student beats teacher</div></div>
	<div class="bs-item"><div class="bs-num">Δ = 0.5333</div><div class="bs-label">PGR slope<br>theory = fitted</div></div>
	</div>
	</section>

	<div class="body-grid" data-measure-role="body">
	<div class="column" data-measure-role="column">
	<div class="card highlight" data-measure-role="card">
	<div class="section-title"><span class="num">1</span><span class="st-text">Spectral transfer, mode by mode</span></div>
	<p class="body-text">Diagonalizing the student dynamics turns distillation into a filter. Modes within the student's horizon copy the teacher; unresolved tail modes are strongly attenuated.</p>
	<div class="eqn"><span class="label">Theorem 1 anatomy</span>
	$$R = \\underbrace{{E_T}}{{\\text{{teacher error}}}} + \\underbrace{{E_S}}{{\\text{{optimization}}}} + \\underbrace{{B_\\mathrm{{align}}}}{{\\text{{alignment bias}}}}$$
	</div>
	<div class="callout"><strong>Prediction:</strong> one-to-one inheritance in learned modes; $O(\\delta^2)$ inheritance in unlearned modes.</div>
	</div>
	<div class="card" data-measure-role="card" data-logbook-target="{CLAIM_1}">
	<div class="section-title"><span class="num">2</span><span class="st-text">Claim 1 · risk inheritance</span></div>
	<span class="scope-tag">Exact diagonal instantiation</span>
	<div class="figure claim1-large"><img src="assets/claim1.png" alt="Mode-wise teacher error inheritance against student learning factor"><div class="caption">At $\\delta=0.1$, head ≥0.9827 while tail ≤0.009956, satisfying $\\delta^2=0.01$.</div></div>
	<div class="verdict">Supported mechanistically</div>
	</div>
	</div>

	<div class="column" data-measure-role="column">
	<div class="card highlight" data-measure-role="card" data-logbook-target="{CLAIM_2}">
	<div class="section-title"><span class="num">3</span><span class="st-text">Claim 2 · horizon expansion</span></div>
	<span class="scope-tag">Analytic proxy · exponent check</span>
	<div class="eqn"><span class="label">Theorem 3</span>
	$$\\mathrm{{DER}}=\\widetilde\\Omega(N^\\kappa),\\quad \\kappa=(\\alpha_T-1-\\beta)(1/\\alpha_T-1/\\alpha_S)=1/6$$
	</div>
	<div class="figure tight"><img src="assets/claim2.png" alt="DER and spectral horizon ratio as sample size grows"><div class="caption">Proxy slope is 0.166667 and DER reaches 46.4 at $N=10^{{10}}$; this checks the exponent only.</div></div>
	</div>
	<div class="card" data-measure-role="card" data-logbook-target="{CLAIM_4}">
	<div class="section-title"><span class="num">4</span><span class="st-text">Claim 4 · spectral denoising</span></div>
	<div class="eqn"><span class="label">Theorem 5 / Eq. 7</span>
	$$1-\\mathrm{{PGR}}=\\widetilde O\\!\\left((k^\\dagger)^{{\\frac{{2\\alpha_S}}{{2\\alpha_S+1}}}}N^{{-\\Delta}}\\right),\\quad \\Delta=0.5333$$
	</div>
	<div class="figure tight"><img src="assets/claim4.png" alt="Performance gap recovered converges to one"><div class="caption">Fitted $\\Delta=0.533333$; PGR = 0.999952 at the largest grid point.</div></div>
	</div>
	</div>

	<div class="column" data-measure-role="column">
	<div class="card highlight" data-measure-role="card" data-logbook-target="{CLAIM_3}">
	<div class="section-title"><span class="num">5</span><span class="st-text">Claim 3 · weak → strong</span></div>
	<span class="scope-tag">Paper-scale synthetic dimensions · 64 seeds</span>
	<p class="body-text">With $d=100$, $N=2000$ and $k^\\dagger\\in\\{{1,10,20,30,50\\}}$, the selected strong student beats its weak teacher and does so in all five tested signal settings.</p>
	<table class="result-table">
	<thead><tr><th class="method">$k^\\dagger$</th><th>1</th><th>10</th><th>20</th><th>30</th><th>50</th></tr></thead>
	<tbody><tr class="ours"><td class="method">best risk / teacher</td><td>0.145</td><td>0.524</td><td>0.714</td><td>0.818</td><td>0.928</td></tr>
	<tr><td class="method">best step</td><td>11</td><td>51</td><td>80</td><td>115</td><td>182</td></tr></tbody>
	</table>
	<div class="callout emph mt-3"><strong>W2S observed:</strong> all five selected students surpass the teacher when intrinsic signal dimension is restricted.</div>
	</div>
	<div class="card" data-measure-role="card" data-logbook-target="{CLAIM_5}">
	<div class="section-title"><span class="num">6</span><span class="st-text">Claim 5 · stop before noise returns</span></div>
	<div class="figure early-large"><img src="assets/claim3.png" alt="Student risk over training steps for five signal dimensions"><div class="caption">Every selected checkpoint beats the final checkpoint. Early stopping is beneficial across all five settings in this synthetic replication.</div></div>
	<div class="verdict">Synthetic mechanism reproduced</div>
	</div>
	</div>

	<div class="column" data-measure-role="column">
	<div class="card highlight" data-measure-role="card" data-logbook-target="{CLAIM_5}">
	<div class="section-title"><span class="num">7</span><span class="st-text">Architecture spectra</span></div>
	<span class="scope-tag">Scaled Figure 2(c) proxy</span>
	<div class="figure arch-large"><img src="assets/architecture.png" alt="Log-log feature covariance spectra for three pretrained architectures on UTKFace"><div class="caption">1,000 UTKFace images; top-500 covariance eigenvalues. Fitted $\\alpha$: ResNet18 1.986, ResNet50 1.551, ViT-B/16 1.628.</div></div>
	<p class="body-text">Log–log fit $R^2$ ranges from 0.950 to 0.989, supporting approximate power-law decay in this scaled sample. This does not test downstream age-regression performance.</p>
	</div>
	<div class="card" data-measure-role="card">
	<div class="section-title"><span class="num">8</span><span class="st-text">Audit boundary</span></div>
	<table class="result-table">
	<thead><tr><th class="method">Evidence</th><th>Scale</th><th>Verdict</th></tr></thead>
	<tbody>
	<tr class="ours"><td class="method">Claims 1–4</td><td>exact / synthetic</td><td>supported</td></tr>
	<tr><td class="method">Figure 2 spectra</td><td>1k images, 3 nets</td><td>scaled support</td></tr>
	<tr><td class="method">Full Figures 2–3</td><td>not rerun</td><td>partial</td></tr>
	</tbody>
	</table>
	<div class="audit-note mt-3">The paper source links no official code. This independent audit publishes scripts, raw CSVs, jobs and seeds.</div>
	</div>
	</div>
	</div>

	<section class="takeaways-strip" data-measure-role="footer-strip">
	<div class="ts-title"><span class="num">9</span> Evidence ladder</div>
	<div class="ts-item"><span class="ts-key">Inheritance</span><span class="ts-text">head ≈ 1; tail ≤ δ²</span></div>
	<div class="ts-item"><span class="ts-key">Expansion</span><span class="ts-text">DER ∝ N¹⁄⁶</span></div>
	<div class="ts-item"><span class="ts-key">Denoising</span><span class="ts-text">PGR → 1</span></div>
	<div class="ts-item"><span class="ts-key">Practice</span><span class="ts-text">selected beats final in 5/5</span></div>
	</section>

	<div class="footer" data-measure-role="footer">
	<div><strong class="method-name">INDEPENDENT REPRODUCTION <span data-ps-mark="woven" data-color-exempt="logo" aria-hidden="true"><svg viewBox="0 0 100 100"><use href="#psReg"/></svg></span></strong> · Tesla T4 · NumPy/PyTorch · 2026-07-21</div>
	<div>Paper, commands, raw results and poster: <span class="repo">Hugging Face Trackio logbook</span></div>
	</div>
	<span class="corner-sig" data-ps-mark="corner" data-color-exempt="logo" aria-hidden="true"><svg viewBox="0 0 100 100"><use href="#psReg"/></svg></span>
	</div>
	</body>
	</html>
	"""


	def main() -> None:
	text = TEMPLATE.read_text(encoding="utf-8")
	text = text.replace("<!DOCTYPE html>", "<!DOCTYPE html>\n" + LOCK, 1)
	swaps = {
	"#2D5F8B": "#315D8A", "#1F4566": "#1D3F60", "#E8F1F8": "#E7F0F8",
	"#D7E5F0": "#D2E2F0", "#C9A24A": "#B45F22", "#FFF7E0": "#FFF0E5",
	"#14314A": "#38200F", "#F6F2F0": "#F4F1ED",
	}
	for old, new in swaps.items():
	text = text.replace(old, new)
	text = text.replace("<title>POSTER TITLE — Venue Year</title>", "<title>What Makes a Strong Model? — ICML 2026 Reproduction</title>")
	text = text.replace("</style>", CUSTOM_CSS + "\n</style>", 1)
	text = text[:text.index("<body>")] + BODY
	OUT.parent.mkdir(parents=True, exist_ok=True)
	OUT.write_text(text, encoding="utf-8")


	if __name__ == "__main__":
	main()

Xet Storage Details

Size:: 14.1 kB
Xet hash:: 652b8b3366dc6267fbe7bf7cb9c3ae00fd24cd7da00fd54dd7227d5115564253

Xet efficiently stores files, intelligently splitting them into unique chunks and accelerating uploads and downloads. More info.