asingh15's picture
Update FineProofs rollout distribution explorer (batch 1/1)
de181bd verified
Raw
History Blame Contribute Delete
14.1 kB
<!doctype html>
<html lang="en">
<head>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1">
<meta name="description" content="Explore 3.31 million continuation grades from the completed FineProofs branch rollout run.">
<title>FineProofs Rollout Distribution</title>
<link rel="preconnect" href="https://cdn.jsdelivr.net">
<link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/katex@0.16.11/dist/katex.min.css" crossorigin="anonymous">
<link rel="stylesheet" href="style.css">
</head>
<body>
<a class="skip-link" href="#main-content">Skip to content</a>
<header class="app-header">
<div class="page-width header-inner">
<div class="brand-block">
<i data-lucide="chart-spline" aria-hidden="true"></i>
<div>
<h1>FineProofs Rollout Distribution</h1>
<p id="run-label">Completed branch collection</p>
</div>
</div>
<a class="dataset-link" href="https://huggingface.co/datasets/asingh15/fineproofs-branch-rollout-distribution" target="_blank" rel="noreferrer">
<i data-lucide="database" aria-hidden="true"></i>
Dataset
</a>
</div>
</header>
<section class="stat-band" aria-label="Collection statistics">
<div class="page-width stat-grid">
<div class="stat"><span>Problems</span><strong id="stat-problems">-</strong></div>
<div class="stat"><span>Base traces</span><strong id="stat-traces">-</strong></div>
<div class="stat"><span>Prefix rows</span><strong id="stat-prefixes">-</strong></div>
<div class="stat"><span>Grades</span><strong id="stat-grades">-</strong></div>
<div class="stat"><span>Mean grade</span><strong id="stat-mean">-</strong></div>
</div>
</section>
<nav class="view-nav" aria-label="Explorer views">
<div class="page-width view-tabs" role="tablist">
<button type="button" role="tab" aria-selected="true" aria-controls="overview-view" id="overview-tab" data-view="overview">
<i data-lucide="layout-dashboard" aria-hidden="true"></i>
Overview
</button>
<button type="button" role="tab" aria-selected="false" aria-controls="trajectories-view" id="trajectories-tab" data-view="trajectories" tabindex="-1">
<i data-lucide="git-branch" aria-hidden="true"></i>
Trajectories
</button>
<button type="button" role="tab" aria-selected="false" aria-controls="method-view" id="method-tab" data-view="method" tabindex="-1">
<i data-lucide="book-open" aria-hidden="true"></i>
Method
</button>
</div>
</nav>
<main id="main-content">
<section class="view page-width" id="overview-view" role="tabpanel" aria-labelledby="overview-tab">
<div class="section-heading">
<div>
<p class="eyebrow">Complete collection</p>
<h2>Grade and value distributions</h2>
</div>
<div class="definition-note">
<strong>Prefix value</strong>
<span>Mean normalized rubric grade over 32 continuations</span>
</div>
</div>
<div class="overview-grid">
<section class="chart-panel wide" aria-labelledby="continuation-title">
<div class="panel-heading">
<div>
<h3 id="continuation-title">Individual continuation grades</h3>
<p id="continuation-summary">All normalized rubric scores</p>
</div>
<span class="panel-badge">n = 3,312,576</span>
</div>
<div class="chart-frame"><svg id="continuation-chart" viewBox="0 0 760 300" role="img" aria-label="Histogram of individual continuation grades"></svg></div>
</section>
<section class="chart-panel" aria-labelledby="credit-title">
<div class="panel-heading">
<div>
<h3 id="credit-title">Credit composition</h3>
<p>Exact zero, partial, and full credit</p>
</div>
</div>
<div id="credit-composition" class="composition"></div>
</section>
<section class="chart-panel wide" aria-labelledby="prefix-title">
<div class="panel-heading">
<div>
<h3 id="prefix-title">Prefix mean grades</h3>
<p>One empirical value per collected prefix</p>
</div>
<span class="panel-badge" id="prefix-median">Median -</span>
</div>
<div class="chart-frame"><svg id="prefix-chart" viewBox="0 0 760 300" role="img" aria-label="Histogram of mean continuation grade by prefix"></svg></div>
</section>
<section class="chart-panel full" aria-labelledby="depth-title">
<div class="panel-heading depth-heading">
<div>
<h3 id="depth-title">Value across reasoning depth</h3>
<p>Mean and interquartile range within normalized-depth deciles</p>
</div>
<div class="legend" aria-label="Endpoint groups">
<span><i class="legend-line all"></i>All traces</span>
<span><i class="legend-line correct"></i>Endpoint grade at least 0.5</span>
<span><i class="legend-line incorrect"></i>Endpoint grade below 0.5</span>
</div>
</div>
<div class="chart-frame depth-frame"><svg id="depth-chart" viewBox="0 0 1080 340" role="img" aria-label="Prefix value by normalized reasoning depth and endpoint grade"></svg></div>
</section>
</div>
</section>
<section class="view trajectory-view" id="trajectories-view" role="tabpanel" aria-labelledby="trajectories-tab" hidden>
<aside class="trajectory-sidebar">
<div class="sidebar-heading">
<div><p class="eyebrow">All base traces</p><h2>Trajectories</h2></div>
<output id="trajectory-count">-</output>
</div>
<form class="filters" id="trajectory-filters">
<label class="search-control">
<i data-lucide="search" aria-hidden="true"></i>
<span class="sr-only">Search problem or ID</span>
<input id="search-input" type="search" placeholder="Search problem or ID" autocomplete="off">
</label>
<div class="filter-grid">
<label><span>Pattern</span><select id="pattern-filter"><option value="all">All patterns</option><option value="rising">Rising</option><option value="falling">Falling</option><option value="volatile">Volatile</option><option value="flat">Flat</option></select></label>
<label><span>Endpoint</span><select id="endpoint-filter"><option value="all">All grades</option><option value="correct">At least 0.5</option><option value="incorrect">Below 0.5</option></select></label>
</div>
<button class="reset-button" id="reset-filters" type="reset"><i data-lucide="rotate-ccw" aria-hidden="true"></i>Reset</button>
</form>
<div class="trajectory-list" id="trajectory-list" role="listbox" aria-label="Filtered trajectories"></div>
</aside>
<section class="trajectory-workspace">
<div class="trajectory-title-row">
<div>
<p class="eyebrow" id="trajectory-position">Trajectory</p>
<h2 id="trajectory-title">-</h2>
<p id="trajectory-meta">-</p>
</div>
<div id="trace-switch" class="trace-switch" role="group" aria-label="Base trace"></div>
</div>
<section class="problem-section" aria-labelledby="problem-title">
<h3 id="problem-title">Problem</h3>
<div id="problem-text" class="problem-text math-content"></div>
</section>
<section class="trajectory-chart-section" aria-labelledby="trajectory-chart-title">
<div class="panel-heading">
<div><h3 id="trajectory-chart-title">Mean grade by prefix</h3><p>Empirical quartiles across 32 continuations; diamond marks the base endpoint grade</p></div>
<span class="pattern-badge" id="pattern-badge">-</span>
</div>
<div class="chart-frame trajectory-chart-frame"><svg id="trajectory-chart" viewBox="0 0 940 350" role="group" aria-label="Selected trajectory prefix values"></svg></div>
</section>
<section class="reasoning-section" id="reasoning-section" aria-labelledby="reasoning-title" aria-busy="false">
<div class="reasoning-heading">
<div>
<p class="eyebrow">Base model output</p>
<h3 id="reasoning-title">Reasoning trace</h3>
<p id="reasoning-context">Selected prefix</p>
</div>
<div class="reasoning-modes" role="group" aria-label="Reasoning trace range">
<button type="button" data-reasoning-mode="prefix" aria-pressed="true">Selected prefix</button>
<button type="button" data-reasoning-mode="segment" aria-pressed="false">New segment</button>
<button type="button" data-reasoning-mode="full" aria-pressed="false">Full trace</button>
</div>
</div>
<dl class="reasoning-meta" aria-label="Reasoning selection metadata">
<div><dt>Range</dt><dd id="reasoning-range">-</dd></div>
<div><dt>Steps</dt><dd id="reasoning-steps">-</dd></div>
<div><dt>Characters</dt><dd id="reasoning-chars">-</dd></div>
</dl>
<div class="reasoning-body">
<div class="reasoning-state" id="reasoning-state" role="status" aria-live="polite">
<span class="reasoning-state-indicator" aria-hidden="true"></span>
<span id="reasoning-state-text">Loading reasoning trace</span>
</div>
<div class="reasoning-empty" id="reasoning-empty" hidden>This prefix contains no reasoning text.</div>
<pre class="reasoning-text" id="reasoning-text" tabindex="0" aria-label="Reasoning trace text" hidden></pre>
<div class="reasoning-tail" id="reasoning-tail" hidden>
<div class="reasoning-tail-label">
<strong>Ungraded tail</strong>
<span id="reasoning-tail-range">Not a collected prefix</span>
</div>
<pre class="reasoning-text reasoning-tail-text" id="reasoning-tail-text" tabindex="0" aria-label="Ungraded reasoning tail"></pre>
</div>
</div>
</section>
<section class="point-section" aria-labelledby="point-title">
<div class="point-heading">
<div><p class="eyebrow">Selected prefix</p><h3 id="point-title">-</h3></div>
<div class="point-controls">
<button class="icon-button" id="previous-point" type="button" title="Previous prefix"><i data-lucide="chevron-left" aria-hidden="true"></i><span class="sr-only">Previous prefix</span></button>
<button class="icon-button" id="next-point" type="button" title="Next prefix"><i data-lucide="chevron-right" aria-hidden="true"></i><span class="sr-only">Next prefix</span></button>
</div>
</div>
<dl class="point-stats" id="point-stats"></dl>
<div class="point-distribution">
<div class="subheading"><h4>Continuation grade distribution</h4><span>32 samples</span></div>
<div class="chart-frame compact-chart"><svg id="point-chart" viewBox="0 0 760 230" role="img" aria-label="Grade histogram for the selected prefix"></svg></div>
</div>
</section>
</section>
</section>
<section class="view page-width method-view" id="method-view" role="tabpanel" aria-labelledby="method-tab" hidden>
<div class="section-heading"><div><p class="eyebrow">Run semantics</p><h2>How to read the collection</h2></div></div>
<div class="method-grid">
<section><span class="method-index">01</span><h3>Prefix value</h3><p>Every partial point is the arithmetic mean of 32 normalized rubric grades from fresh continuations of that exact response prefix. It is not a binary success probability.</p></section>
<section><span class="method-index">02</span><h3>Trajectory endpoint</h3><p>The diamond is the original base response's continuous rubric grade. The downstream PRM dataset separately thresholds eligible endpoints at 0.5 for a hard yes/no label.</p></section>
<section><span class="method-index">03</span><h3>Empty prefix</h3><p>Each problem contributes two selected base traces, so its empty prefix appears once per trace in the raw branch set. The 103,518 rows correspond to 99,354 unique semantic problem-prefix inputs.</p></section>
<section><span class="method-index">04</span><h3>PRM filtering</h3><p id="prm-accounting">The downstream training build removes duplicate empty prefixes, invalid endpoints, and overlength prompts while preserving provenance counts.</p></section>
</div>
<dl class="run-details" id="run-details"></dl>
</section>
</main>
<footer>
<div class="page-width footer-inner">
<span>Qwen3.5-9B continuations · GPT-OSS-120B rubric grades</span>
<code id="run-id">-</code>
</div>
</footer>
<script defer src="https://cdn.jsdelivr.net/npm/lucide@0.468.0/dist/umd/lucide.min.js"></script>
<script defer src="https://cdn.jsdelivr.net/npm/katex@0.16.11/dist/katex.min.js" crossorigin="anonymous"></script>
<script defer src="https://cdn.jsdelivr.net/npm/katex@0.16.11/dist/contrib/auto-render.min.js" crossorigin="anonymous"></script>
<!-- DATA_SCRIPTS_START -->
<script defer src="data-0.js"></script>
<script defer src="data-1.js"></script>
<script defer src="data-2.js"></script>
<!-- DATA_SCRIPTS_END -->
<script defer src="app.js"></script>
</body>
</html>