index.html

<!DOCTYPE html>
<html lang="en">
<head>
<meta charset="utf-8">
<meta name="viewport" content="width=device-width, initial-scale=1">
<title>ColdSlither Labs — Local LLM Research and Build</title>
<style>
:root{--bg:#0d0f12;--card:#161a20;--ink:#e8e6e1;--mut:#9aa3ad;--acc:#e0a458;--line:#262c34}
*{box-sizing:border-box}body{margin:0;background:var(--bg);color:var(--ink);font:16px/1.6 system-ui,-apple-system,Segoe UI,Roboto,sans-serif}
.wrap{max-width:820px;margin:0 auto;padding:40px 20px 64px}
.hero{border:1px solid var(--line);border-radius:14px;padding:32px;background:linear-gradient(180deg,#141821,#0d0f12)}
.hero h1{margin:0 0 6px;font-size:30px;letter-spacing:.5px}
.hero h1 span{color:var(--acc)}
.sub{color:var(--mut);margin:0 0 18px}
.badges{display:flex;flex-wrap:wrap;gap:8px;margin:0 0 6px}
.badges b{font-weight:600;font-size:13px;border:1px solid var(--line);border-radius:20px;padding:4px 12px;color:var(--mut)}
.grid{display:grid;grid-template-columns:1fr 1fr;gap:14px;margin:22px 0}
@media(max-width:640px){.grid{grid-template-columns:1fr}}
.card{background:var(--card);border:1px solid var(--line);border-radius:12px;padding:18px}
.card h2{margin:0 0 8px;font-size:17px;color:var(--acc)}
.card p,.card li{font-size:14.5px;color:var(--ink)}
.card ul{margin:8px 0 0;padding-left:18px}
.num{font-size:26px;font-weight:700;color:var(--acc)}
.caveat{font-size:13px;color:var(--mut)}
footer{margin-top:26px;color:var(--mut);font-size:14px}
a{color:var(--acc)}
</style>
</head>
<body>
<div class="wrap">
<div class="hero">
<h1>ColdSlither <span>Labs</span></h1>
<p class="sub">Local LLM research and build. Small models, honest numbers, reproducible recipes. Based in East Orange, NJ. Working toward technical independence, one verified run at a time.</p>
<div class="badges"><b>QLoRA fine-tuning</b><b>Refusal research</b><b>Quantization</b><b>Eval discipline</b><b>RTX 5060 Ti 16GB</b></div>
</div>
<div class="grid">
<div class="card"><h2>Capability-preserving uncensoring</h2>
<p><span class="num">0 refusals</span><br>OBLITERATUS advanced run on Qwen3-8B with capability intact, served as GGUF. Selection rule: lowest KL divergence under 10 refusals, never min refusal. KL over 0.5 means real damage.</p>
<p class="caveat">Measured on our prompt set and HF 0/32; harsher shared harness scores 4/10. Both numbers published, neither hidden.</p></div>
<div class="card"><h2>Head-to-head method comparison</h2>
<p>Heretic vs OBLITERATUS on the same Qwen3-8B base. Both held 5/5 capability. Refusal: Heretic 30%, OBLITERATUS 40% (n=10). Heretic is the locked default; OBLITERATUS stays as manual fallback.</p></div>
<div class="card"><h2>Anti-collapse training module</h2>
<p><span class="num">7.72 held-out</span><br>Arenas residual module, trained and benchmarked on Qwen3-0.6B/1.7B. Beats the naive no-residual foil in every run. Magnitude-normalized residual leads.</p></div>
<div class="card"><h2>Eval discipline</h2>
<p>Content-only refusal scoring, 400 to 600 token budgets, matched-budget ablations. A 2x2 isolation proved an apparent serving gap was a token-budget artifact, not a thinking-mode effect. Measure twice, publish once.</p></div>
</div>
<div class="card"><h2>What I offer</h2>
<ul>
<li>Local model setup on consumer GPUs: serve open-weight models privately, no cloud dependency.</li>
<li>Fine-tuning guidance: QLoRA recipes that fit 16GB cards, resumable runs, honest evals.</li>
<li>Security-minded eval: refusal boundary probing and regression checks for teams shipping local models.</li>
</ul>
<p class="caveat">Research stage. Methods are measured on small harnesses and published with caveats. No enterprise claims, no borrowed logos.</p></div>
<footer>Contact via <a href="https://huggingface.co/ColdSlither">huggingface.co/ColdSlither</a>. Every number above ships with its harness, or it does not ship.</footer>
</div>
</body>
</html>