README.md

metadata

title: Tablinum
emoji: 🏛
colorFrom: blue
colorTo: yellow
sdk: static
pinned: false

Tablinum

the record room added to a house that already existed

Small models that do real work on small devices.

We take an existing base model and attach what it needs — trained office skills, a fact-table sidecar, a deterministic arithmetic harness.

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.

The Dose-Window Hypothesis — the paper behind these claims: order, calibration, and externalised knowledge in small language models.

354M

smallest production coworker

1.000

eval.v1 — stock bases sit at ≈0.24

0.980

m365 premise bench · sidecar, single 8B

7 rungs measured

every quant we ship, scored

The models

Two coworkers, one contract

cowork-companion

The default: 354M, one JSON envelope, decks / sheets / documents / grounded refusal. Runs on a laptop at Q8_0.

Q8_0 · 379 MB0.730

Q6_K · 293 MB0.695

Q5_K_M · 260 MB0.708

Q4_K_M · 229 MB0.517 — the cliff

cowork-companion-1.2b

The quality-max sibling: same recipe, 3.4× parameters — bought for full-document docx renovation on real files.

Q8_0 · 1.25 GB0.919

Q6_K · 963 MB0.924

Q5_K_M · 843 MB0.918

The method

Three measured claims

I.

The skills are trained, not bought.

Stock bases up to 3.4× larger are near-chance (≈0.24) on our office instruments; the trained 354M scores 1.000.

II.

Knowledge comes from the sidecar.

Tabula — the fact tablet you attach — lifts premise correction 4–5× at zero training and carries a single 8B to 0.980 on a blind holdout.

III.

Arithmetic goes to a harness.

One-shot multi-row totals fail at every scale we measured; the --compute ledger answers exactly regardless of model state.

Reading the numbers

Every score recomputes

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.

The set ships complete: the 3,258-row training table, both models at every measured rung, and the open-source cowork CLI (constrained decode, exact --compute arithmetic, artifact rendering, the Tabula sidecar). Instrument hashes are published; instrument contents stay frozen and unpublished.

Tablinum — the room; Tabula — the tablet in it.