| ---
|
| title: Tablinum
|
| emoji: "🏛"
|
| colorFrom: blue
|
| colorTo: yellow
|
| sdk: static
|
| pinned: false
|
| ---
|
|
|
| <style>
|
| :root{--cinnabar:#C8392E;--cinnabar-soft:#D86F65;--bg-blue:#1E4A7E;--bg-blue-2:#2E5A8E;
|
| --tyrian:#5D1B6A;--gold:#C9A862;--gold-deep:#A07B14;--orpiment:#E8B547;--slate:#6E5236;
|
| --slate-deep:#3D2A1B;--celadonite:#7A8A4A;--lamp:#1A1612;--linen:#F4E4C9;--linen-2:#ECE4D2;
|
| --paper:#FBF4DF;--rule:#C8A87E;--rule-soft:#E0D2B4;
|
| --serif:Georgia,'Iowan Old Style','Palatino Linotype',serif;
|
| --sans:ui-sans-serif,system-ui,'Segoe UI',Roboto,sans-serif;
|
| --mono:ui-monospace,'Cascadia Mono','Segoe UI Mono',Consolas,monospace}
|
| .tlin{max-width:980px;margin:0 auto;background:var(--linen);color:var(--lamp);
|
| font-family:var(--sans);font-size:15px;line-height:1.6;border:1px solid var(--rule);
|
| box-shadow:0 1px 0 rgba(110,82,54,.1)}
|
| .tlin *{box-sizing:border-box}
|
| .tlin .strip{height:108px;background:linear-gradient(135deg,var(--bg-blue) 0%,var(--bg-blue-2) 100%);
|
| display:flex;align-items:center;gap:18px;padding:0 28px;position:relative;overflow:hidden}
|
| .tlin .strip::after{content:"";position:absolute;inset:0;
|
| background-image:radial-gradient(circle at 20% 50%,rgba(255,255,255,.05) 0 1px,transparent 1.5px);
|
| background-size:11px 11px;pointer-events:none;opacity:.7}
|
| .tlin .mark{width:64px;height:64px;background:rgba(251,244,223,.95);border-radius:7px;
|
| display:flex;align-items:center;justify-content:center;box-shadow:0 3px 14px rgba(0,0,0,.35);z-index:1}
|
| .tlin .mark svg{width:44px;height:44px}
|
| .tlin .wordmark{z-index:1}
|
| .tlin .wordmark .org{font-size:13px;letter-spacing:.24em;text-transform:uppercase;
|
| color:var(--gold);font-weight:600}
|
| .tlin .wordmark .full{font-family:var(--serif);font-size:17px;color:var(--linen);
|
| letter-spacing:.04em;margin-top:2px}
|
| .tlin .hero{padding:22px 28px 16px}
|
| .tlin .hero .name{font-family:var(--serif);font-size:30px;font-weight:600;margin:0}
|
| .tlin .hero .tag{font-family:var(--serif);font-style:italic;font-size:14.5px;color:var(--slate-deep);margin:4px 0 0}
|
| .tlin .hero .crule{height:2px;background:var(--cinnabar);width:60px;margin:12px 0 10px}
|
| .tlin .hero p{margin:0;max-width:66ch}
|
| .tlin .metrics{display:grid;grid-template-columns:repeat(4,1fr);margin:0 28px 8px;
|
| background:var(--linen-2);border:1px solid var(--rule);border-radius:8px;overflow:hidden}
|
| .tlin .metric{padding:14px;border-right:1px solid var(--rule-soft)}
|
| .tlin .metric:last-child{border-right:0}
|
| .tlin .metric .v{font-family:var(--serif);font-size:23px;font-weight:600;line-height:1.1}
|
| .tlin .metric .v.cin{color:var(--cinnabar)}.tlin .metric .v.gold{color:var(--gold-deep)}
|
| .tlin .metric .v.blue{color:var(--bg-blue)}.tlin .metric .v.cel{color:var(--celadonite);font-size:15px;padding-top:5px}
|
| .tlin .metric .l{font-family:var(--mono);font-size:10px;letter-spacing:.06em;color:var(--slate);
|
| text-transform:uppercase;margin-top:3px}
|
| .tlin section{padding:26px 28px;border-top:1px solid var(--rule-soft)}
|
| .tlin section:first-of-type{border-top:0}
|
| .tlin .k{font-family:var(--mono);font-size:10.5px;letter-spacing:.18em;text-transform:uppercase;
|
| color:var(--gold-deep);font-weight:600;margin:0 0 6px}
|
| .tlin h2{font-family:var(--serif);font-size:23px;font-weight:600;margin:0 0 10px}
|
| .tlin .models{display:grid;grid-template-columns:1fr 1fr;gap:14px}
|
| .tlin .model{background:var(--paper);border:1px solid var(--rule);border-radius:8px;padding:16px}
|
| .tlin .model .nm{font-family:var(--serif);font-size:19px;font-weight:600;margin:0}
|
| .tlin .model .nm a{color:var(--bg-blue);text-decoration:none;border-bottom:1px solid var(--rule)}
|
| .tlin .model .ds{font-size:13px;color:var(--slate-deep);margin:6px 0 10px}
|
| .tlin .rung{display:flex;justify-content:space-between;font-family:var(--mono);font-size:11.5px;
|
| padding:4px 0;border-top:1px dashed var(--rule-soft)}
|
| .tlin .rung b{font-weight:600}
|
| .tlin .rung .good{color:var(--celadonite);font-weight:700}
|
| .tlin .rung .mid{color:var(--gold-deep);font-weight:700}
|
| .tlin .rung .bad{color:var(--cinnabar);font-weight:700}
|
| .tlin .claims{display:grid;grid-template-columns:1fr 1fr 1fr;gap:14px}
|
| .tlin .claim{background:var(--paper);border:1px solid var(--rule);border-radius:8px;padding:14px}
|
| .tlin .claim .no{font-family:var(--serif);font-size:15px;color:var(--cinnabar);font-weight:700}
|
| .tlin .claim .tt{font-family:var(--serif);font-size:15.5px;font-weight:600;margin:3px 0 5px}
|
| .tlin .claim .tx{font-size:12.8px;color:var(--slate-deep)}
|
| .tlin .claim .tx b{color:var(--lamp)}
|
| .tlin .foot{padding:18px 28px 22px;color:var(--slate);font-size:12.5px;font-style:italic;
|
| font-family:var(--serif)}
|
| @media (max-width:760px){.tlin .metrics{grid-template-columns:1fr 1fr}
|
| .tlin .models,.tlin .claims{grid-template-columns:1fr}}
|
| </style>
|
|
|
| <div class="tlin">
|
| <div class="strip">
|
| <div class="mark"><img src="https://huggingface.co/spaces/tablinum/README/resolve/main/logo.png" alt="Tablinum mark" style="width:54px;height:54px;border-radius:8px;"></div>
|
| <div class="wordmark"><div class="org">Tablinum</div>
|
| <div class="full">the record room added to a house that already existed</div></div>
|
| </div>
|
| <div class="hero">
|
| <h1 class="name">Small models that do real work on small devices.</h1>
|
| <p class="tag">We take an existing base model and attach what it needs — trained office
|
| skills, a fact-table sidecar, a deterministic arithmetic harness.</p>
|
| <div class="crule"></div>
|
| <p>Nothing here is a general assistant. Every model is a measured instrument,
|
| published at every quantization rung we ship, with each rung's score on the
|
| same frozen instrument. When a rung degrades, the card says so.</p>
|
| <p style="margin-top:10px"><a href="https://huggingface.co/spaces/tablinum/dose-window-hypothesis"
|
| style="color:var(--bg-blue);font-weight:600">The Dose-Window Hypothesis</a>
|
| — the paper behind these claims: order, calibration, and externalised
|
| knowledge in small language models.</p>
|
| </div>
|
| <div class="metrics">
|
| <div class="metric"><div class="v gold">354M</div><div class="l">smallest production coworker</div></div>
|
| <div class="metric"><div class="v cin">1.000</div><div class="l">eval.v1 — stock bases sit at ≈0.24</div></div>
|
| <div class="metric"><div class="v blue">0.980</div><div class="l">m365 premise bench · sidecar, single 8B</div></div>
|
| <div class="metric"><div class="v cel">7 rungs measured</div><div class="l">every quant we ship, scored</div></div>
|
| </div>
|
|
|
| <section>
|
| <p class="k">The models</p>
|
| <h2>Two coworkers, one contract</h2>
|
| <div class="models">
|
| <div class="model">
|
| <p class="nm"><a href="https://huggingface.co/tablinum/cowork-companion">cowork-companion</a></p>
|
| <p class="ds">The default: 354M, one JSON envelope, decks / sheets / documents /
|
| grounded refusal. Runs on a laptop at Q8_0.</p>
|
| <div class="rung"><span><b>Q8_0</b> · 379 MB</span><span class="good">0.730</span></div>
|
| <div class="rung"><span><b>Q6_K</b> · 293 MB</span><span class="mid">0.695</span></div>
|
| <div class="rung"><span><b>Q5_K_M</b> · 260 MB</span><span class="mid">0.708</span></div>
|
| <div class="rung"><span><b>Q4_K_M</b> · 229 MB</span><span class="bad">0.517 — the cliff</span></div>
|
| </div>
|
| <div class="model">
|
| <p class="nm"><a href="https://huggingface.co/tablinum/cowork-companion-1.2b">cowork-companion-1.2b</a></p>
|
| <p class="ds">The quality-max sibling: same recipe, 3.4× parameters — bought for
|
| full-document docx renovation on real files.</p>
|
| <div class="rung"><span><b>Q8_0</b> · 1.25 GB</span><span class="good">0.919</span></div>
|
| <div class="rung"><span><b>Q6_K</b> · 963 MB</span><span class="good">0.924</span></div>
|
| <div class="rung"><span><b>Q5_K_M</b> · 843 MB</span><span class="good">0.918</span></div>
|
| </div>
|
| </div>
|
| </section>
|
|
|
| <section>
|
| <p class="k">The method</p>
|
| <h2>Three measured claims</h2>
|
| <div class="claims">
|
| <div class="claim"><div class="no">I.</div>
|
| <div class="tt">The skills are trained, not bought.</div>
|
| <div class="tx">Stock bases up to <b>3.4× larger are near-chance</b> (≈0.24) on our
|
| office instruments; the trained 354M scores <b>1.000</b>.</div></div>
|
| <div class="claim"><div class="no">II.</div>
|
| <div class="tt">Knowledge comes from the sidecar.</div>
|
| <div class="tx"><b>Tabula</b> — the fact tablet you attach — lifts premise
|
| correction <b>4–5× at zero training</b> and carries a single 8B to
|
| <b>0.980</b> on a blind holdout.</div></div>
|
| <div class="claim"><div class="no">III.</div>
|
| <div class="tt">Arithmetic goes to a harness.</div>
|
| <div class="tx">One-shot multi-row totals fail at <b>every scale we measured</b>;
|
| the <b>--compute</b> ledger answers exactly regardless of model
|
| state.</div></div>
|
| </div>
|
| </section>
|
|
|
| <section>
|
| <p class="k">Reading the numbers</p>
|
| <h2>Every score recomputes</h2>
|
| <p style="max-width:72ch">Each figure on every card comes from a frozen, committed
|
| instrument and saved generations — oracle-graded, shortcut-checked, with the
|
| failure walls printed next to the wins. The quant ladders are measured through
|
| llama.cpp itself, because deployment truth beats spec sheets.</p>
|
| <p style="max-width:72ch">The set ships complete: the
|
| <a href="https://huggingface.co/datasets/tablinum/cowork-companion-sft">3,258-row training table</a>,
|
| both models at every measured rung, and the open-source
|
| <a href="https://github.com/chadstthomas-svg/cowork-cli">cowork CLI</a>
|
| (constrained decode, exact --compute arithmetic, artifact rendering, the
|
| Tabula sidecar). Instrument <a href="https://huggingface.co/tablinum/cowork-companion/blob/main/INSTRUMENTS.sha256">hashes</a>
|
| are published; instrument contents stay frozen and unpublished.</p>
|
| </section>
|
|
|
| <div class="foot">Tablinum — the room; Tabula — the tablet in it.</div>
|
| </div>
|
| |
