Tablinum-3

The Lite-Beam engram, release 3. A tablinum was the record room of a Roman house: the place the household's records were kept.

A library that runs on your own disk and that a model reads before it answers. Tablinum-3 holds 34 tables and 210 million passages: Wikipedia, case law, EU and US law, research abstracts, software documentation, sysadmin and developer Q&A, app manuals, web and MCP specifications, primary humanities texts and home-repair know-how. It is built once, searched in milliseconds on a CPU, and shared by every model on the machine. The model gets the three best passages, each tagged with where it came from, and answers from them.

On Moon-Beam 8B (Q4_K_M), Tablinum-3 nearly triples the right answers on AA-Omniscience, from 9.0% to 26.0%, and lifts the index from −56.0 to +4.0: better on 111 of 200 questions, worse on 19. With only the downloadable packs it still reaches 22.5%.

Quick links: What it holds · Packs · Results · Use it · What is not released · Tablinum-2 · terrabeam

What is new in 3

• Ten new tables, about 49M new passages. Most of them are for agents: sysadmin and developer Q&A, app manuals (VS Code, PowerShell, Godot, Blender, LibreOffice, GIMP), web automation and accessibility specs, the Model Context Protocol, OpenAPI definitions, scientific code. Plus home repair, primary humanities texts and humanities full text.
• Route v3. For each question domain, the route says which tables to search. v3 adds EU law to Finance (answer in the top 3 rose from 0.21 to 0.27 on our retrieval probes) and engineering and biology abstracts to STEM (top 10 from 0.33 to 0.42).
• Packs rebuilt around use. A new agent pack, a lite default that now includes the small agent tables, and the rule that nothing a route does not use sits in lite or standard. See Packs.
• Hosted once. The new tables live here; the Tablinum-2 and Engram 1 tables stay where they are. packs.json points each table at the repository that holds it, so nothing is uploaded twice and every download is checked against its SHA-256.

What it holds

shelf	tables	passages	in the packs
Reference	wiki (English Wikipedia), humref (Britannica 1911), oer (open textbooks)	75.3M	all
Humanities	humprim (Gutenberg, Wikisource, Sefaria), papers_humft, papers_hum, philo, sehum	34.2M	sehum only
Health	health (textbooks, PubMed, StatPearls), healthp, fdalabel, cdcj	29.6M	none: health data stays local
Software and agents	software (DevDocs), sesys, sedev, apps, web, scicode, cli	23.4M	all but cli
Law and regulation	law (US case law), eulex, uscode, ecfr, selaw	22.6M	all
Science and engineering	papers_sci, papers_eng, papers_bio, sestem, math	20.3M	sestem only
Finance	finreg (CRS reports), papers_fin, sefin	4.0M	finreg, sefin
Home and everyday	homerepair, psych	0.6M	homerepair
all	34 tables	210.0M	127.7M in 25 tables

New tables are in bold. Per table:

new table	what is in it	passages	tar
sesys	Super User, Ask Ubuntu, Unix & Linux, Server Fault, Apple, Android, Raspberry Pi, networking, security, DevOps, DBA, Vi, Emacs, TeX Q&A	10.14M	5.71 GB
sedev	Stack Overflow threads about developer and agent tooling, plus Software Engineering, Code Review, Game Development, Blender	7.76M	4.16 GB
homerepair	home improvement, car maintenance, gardening, woodworking, cooking, lifehacks, crafts Q&A	0.54M	0.36 GB
scicode	Rosetta Code tasks in common languages, SciPy 1.18 reference	0.27M	0.11 GB
apps	VS Code, PowerShell, Godot, Blender, LibreOffice, GIMP documentation	0.14M	0.06 GB
web	WAI-ARIA and its mappings, ARIA practices, Selenium, Playwright, MCP spec and reference servers, OpenAPI directory	0.09M	0.04 GB
humprim	Gutenberg humanities classes, Wikisource, Sefaria	18.82M	recipe
papers_humft	full text of humanities papers (peS2o)	10.69M	recipe
cli	tldr pages, Arch Wiki, GNU manuals	0.13M	recipe
psych	psychology, interpersonal skills, parenting, health Q&A	0.09M	recipe

Every passage starts with its source and citation (Ask Ubuntu: How do I…, Godot docs: …, 26 U.S.C. §1), so the model can say where an answer came from and BM25 can match exact names, commands and identifiers.

Packs

A pack is a named set of tables that terrabeam downloads in one go. The packs nest: lite is inside standard, and standard is inside full, so moving up a tier downloads only the difference.

pack	download	tables	for
lite (default)	6.6 GB	software, apps, web, scicode, homerepair, Stack Exchange STEM / humanities / finance / law, Britannica 1911, US Code, eCFR	any laptop; everyday questions, code, apps
agent	12.4 GB	software, apps, web, scicode, sedev, sesys	coding and computer-use agents (Claude Code, opencode, Cline through terrabeam mcp)
standard	31.1 GB	lite + case law, EU law, CRS reports, sedev, sesys	a workstation; adds the tables the law and finance routes use
full	76.0 GB	standard + English Wikipedia + open textbooks	everything that can be downloaded
wiki	33.6 GB	English Wikipedia only

How the tiers were drawn:

1. lite holds small tables with broad use. Every table in it is under 2.5 GB, and together they fit a laptop.
2. standard adds the big tables a route uses. Case law, EU law and CRS reports are on the law and finance routes. sedev and sesys are large (about 10 GB together) and serve agents.
3. full adds what no route uses yet. Wikipedia is the fallback for every domain but is 34 GB. The open textbooks did not lift any domain in our tests (Finance, STEM and humanities scored lower with them), so they moved from standard to full.
4. agent is a use-case pack, not a tier: the agent tables without the rest.
5. Health is never in a pack, whatever its license. The text stays local; the recipes are public.

Does it help?

We tested the library on Moon-Beam 8B in its Q4_K_M GGUF, the file most people will run (byte-identical to the published one). The questions are the 200-question dev split of AA-Omniscience: hard factual questions across six fields, where a wrong answer costs a point and "I don't know" costs nothing.

arm	correct	wrong	AA index	vs closed-book: better / worse
closed-book (no library)	9.0%	65.0%	−56.0
Tablinum-3, downloadable packs only	22.5%	24.5%	−2.0	102 / 19
Tablinum-3, full library	26.0%	22.0%	+4.0	111 / 19

The biggest change is in wrong answers. Without the library the model guesses and is wrong on 130 of 200 questions. With it, wrong answers fall to 44: when the passages hold the answer it uses them, and when they don't it more often says it does not know. Both changes are significant (paired sign test, p < 0.0001).

Downloadable packs vs the full library. The full route also searches the recipe-only tables: health textbooks, research abstracts and philosophy. Without them, Health, STEM and Humanities fall back to Wikipedia, which costs 3.5 points of accuracy. Law and Software are fully covered by the packs.

AA index by field (33-34 questions each, so single fields are noisy):

field	closed-book	downloadable packs	full library
Finance	-51.5	-24.2	-15.2
Health	-57.6	-6.1	+0.0
Humanities	-47.1	+0.0	+0.0
Law	-54.5	+9.1	+15.2
STEM	-63.6	-30.3	-21.2
Software	-61.8	+38.2	+44.1

Route v3 vs v2. On this model the new route ties the Tablinum-2 route: v3 was better on 20 questions and v2 on 18, and v2 scored 27.0% / +3.0. The extra Finance and STEM tables raised retrieval hits in our probes but did not move Moon-Beam's answers beyond the noise of 33 questions per field. The library keeps route v3: it scores the same and adds EU law, which is in the packs, to Finance.

How it was run: thinking off, temperature 1.0 / top-p 0.95 / top-k 20, the three best passages per question after reranking, graded by Claude Haiku with the AA-Omniscience rubric. This is the protocol on the Moon-Beam 8B card, where the BF16 model goes from 7.5% to 27.0% with Tablinum-2, so the Q4 loses little. Totals and per-field numbers are in results/moon-beam-8b-q4_k_m_aa-dev.json.

How it works

Retrieval does not depend on the model: a static embedder for the first pass, BM25 for exact terms, then a small cross-encoder reranks the candidates, all on the CPU. That is why one library can serve every model. What a table is worth does depend on the model, so each model can get its own gate: for every question domain, the same questions are answered with and without that domain's tables, and a table stays on only if it helps more often than it hurts. Tablinum-2 introduced the gate; its card has the Qwen3.8-27B gate.

The routes are in routes.json: route_v3 uses the whole library, and route_v3_published is the same route limited to the tables you can download.

Use it

With terrabeam (any GGUF, any OpenAI-compatible app):

terrabeam pack lite         # or agent, standard, full
terrabeam run tablinum/Moon-Beam-8B-A1B-GGUF-Preview

In Python:

pip install faiss-cpu tantivy pyarrow tokenizers safetensors transformers torch huggingface_hub
git clone https://huggingface.co/datasets/tablinum/tablinum-3 && cd tablinum-3
mkdir -p tables && tar -xf tables/apps.tar -C tables     # one table, or all of them
export ENGRAM_ROOT=$PWD

import sys; sys.path.insert(0, "code")
from engram import Engram
eg = Engram(tables=["apps"])
for h in eg.retrieve("How do I change the default terminal profile in VS Code?", k=3, tables=["apps"]):
    print(h["table"], h["text"][:200])

Tables from Tablinum-2 (eulex, sestem, wiki, …) and Engram 1 (software, law, oer) unpack into the same tables/ folder; packs.json lists where each one lives. code/engram_eval.py runs a closed-book or library arm against any OpenAI-compatible server (--url, --route v3 --tables auto), and code/grade.py grades it.

What is released and what is not

The rules we follow, for every table:

1. Released only if every source allows it. A table is published only when each of its sources permits redistributing verbatim excerpts. Each released table's sources and licenses are listed in SOURCES.md, and every passage names its source.
2. Share-alike stays share-alike. The Stack Exchange tables (sesys, sedev, homerepair) are CC BY-SA 4.0.
3. Health text stays local, even when its license would allow publishing (fdalabel and cdcj are public domain and CC0). Medical passages out of context are a safety risk, so these tables are recipe-only.
4. When one source blocks a table, the whole table is recipe-only rather than quietly dropping that source. humprim contains non-commercial Sefaria translations. cli contains GNU manuals with GFDL invariant sections. papers_* and philo are publisher and Stanford text.
5. Every table has a recipe, released or not, so the whole library can be rebuilt for private use.

License

The Tablinum-3 compilation, meaning the table design, routes, gates, indexes, manifests, recipes and code, is licensed under CC BY 4.0 with required attribution to Tablinum. Any use, redistribution or derivative must credit Tablinum and link this dataset.

The passage text in each released table keeps its upstream license (SOURCES.md). Character art: Beam by Tablinum.

Citation

@misc{tablinum3,
  title  = {Tablinum-3: the Lite-Beam engram},
  author = {Tablinum},
  year   = {2026},
  note   = {A local retrieval library for quantized LLMs, 34 tables and 210M passages},
  url    = {https://huggingface.co/datasets/tablinum/tablinum-3}
}