hyperdex-trainer / web /templates /landing.html
GGUFGuy's picture
Duplicate from GGUFGuy/hyperdex-trainer
2a49d9a
Raw History Blame Contribute Delete
5.07 kB
{% extends "base.html" %}
{% block title %}NanoDex — train a language model from scratch{% endblock %}
{% block body %}
<section class="hero">
<span class="tag"><b>{{ n_devices }}</b> {{ 'GPU' if on_gpu else 'CPU' }} worker{{ 's' if n_devices != 1 }} ·
<b>{{ stats.done }}</b> models born here</span>
<h1>Train your first<br><em>decoder-only model</em><br>from scratch.</h1>
<p>Not a fine-tune. Not an adapter. Random weights, a real corpus, and a real
loss curve you watch fall in real time — from a browser tab, on our GPUs.</p>
<div class="cta">
<a class="btn primary lg" target="_top" href="/login?next=/train">Start a run &nbsp;→</a>
<a class="btn lg" href="/explore">Browse the models</a>
</div>
<p class="frame-note">Signing in opens NanoDex full screen — Hugging Face's
login page can't run inside an embedded frame.</p>
</section>
<section style="margin-top:22px">
<div class="grid g3">
<div class="card">
<div class="step-n">1</div>
<h3>Pick a size</h3>
<p class="sub" style="margin:6px 0 0">Four architectures, from 492k to 8.1M
parameters. That's the <em>real</em> total — embeddings included, no asterisk.</p>
</div>
<div class="card">
<div class="step-n">2</div>
<h3>Pick a token budget</h3>
<p class="sub" style="margin:6px 0 0">Anywhere from
{{ min_tokens // 10**6 }}M to <b>1.5 billion</b> tokens of
<b>fineweb-edu</b>. More tokens, sharper model, longer wait.</p>
</div>
<div class="card">
<div class="step-n">3</div>
<h3>Watch it learn</h3>
<p class="sub" style="margin:6px 0 0">Your run joins the queue. Live loss
curve, throughput, logs — then the weights land in your own Hugging Face
account.</p>
</div>
</div>
</section>
<section style="margin-top:44px">
<div class="sect-h">
<div>
<span class="eyebrow">The four sizes</span>
<h2 style="margin-top:5px">Small enough to finish. Big enough to learn something.</h2>
</div>
</div>
<div class="picks">
{% for t in tiers %}
<div class="pick">
<b>{{ t.label }}</b>
<div class="p-n">{{ "{:,}".format(t.params) }} params</div>
<div class="p-d">
{{ t.layers }} layers · {{ t.hidden }} hidden<br>
{{ t.heads }} heads ({{ t.kv_heads }} KV) · FFN {{ t.ffn }}
</div>
</div>
{% endfor %}
</div>
</section>
<section style="margin-top:44px">
<div class="grid g2">
<div class="card pad-lg">
<span class="eyebrow">What you actually get</span>
<h2 style="margin:8px 0 12px;font-size:21px">A model that is genuinely yours</h2>
<p class="sub">Every finished run is pushed to <b>your</b> Hugging Face
namespace: <span class="mono">config.json</span>,
<span class="mono">model.safetensors</span>, the tokenizer, a
<span class="mono">training_run.json</span> with the full recipe, and a
generated model card. Load it with <span class="mono">transformers</span>
like anything else on the Hub.</p>
</div>
<div class="card pad-lg">
<span class="eyebrow">Honest expectations</span>
<h2 style="margin:8px 0 12px;font-size:21px">It will not be an assistant</h2>
<p class="sub">At a few hundred thousand parameters, a model learns word
shapes, common collocations and a little syntax. It will produce
English-looking text with no facts in it. That's the correct outcome —
and watching cross-entropy fall from <b>7.6</b> (uniform noise over 2,048
tokens) to somewhere near <b>4</b> is the whole point.</p>
</div>
</div>
</section>
{% if recent %}
<section style="margin-top:44px">
<div class="sect-h">
<div>
<span class="eyebrow">Trained here</span>
<h2 style="margin-top:5px">The newest models on the shelf</h2>
</div>
<a class="btn ghost sm" href="/explore">Explore all →</a>
</div>
<div class="runs">
{% for m in recent %}
<a class="run" href="/models/{{ m.id }}">
<div class="run-top">
<h4>{{ m.name }}</h4>
<span class="chip done"><i></i>completed</span>
</div>
<div class="run-meta">
<span class="mono">{{ m.tier_label }}</span>
<span>{{ "{:,}".format(m.n_params) }} params</span>
<span><b>{{ (m.tokens_seen / 1000000)|round|int }}M</b> tokens</span>
{% if m.best_loss %}<span>loss <b>{{ "%.4f"|format(m.best_loss) }}</b></span>{% endif %}
<span>@{{ m.username }}</span>
</div>
</a>
{% endfor %}
</div>
</section>
{% endif %}
<section style="margin-top:44px" class="card pad-lg">
<div style="display:flex;gap:24px;align-items:center;justify-content:space-between;flex-wrap:wrap">
<div>
<h2 style="font-size:21px">Ready when you are.</h2>
<p class="sub" style="margin-top:6px">Sign in with Hugging Face — no separate account, no password.</p>
</div>
<a class="btn primary lg" target="_top" href="/login?next=/train">Start a run &nbsp;→</a>
</div>
</section>
{% endblock %}