Tablinum-2

The Lite-Beam engram, release 2. A tablinum was the record room of a Roman house: the place the household's knowledge was kept.

A local knowledge library for small quantized models: 24 tables and about 160M passages (104M outside Wikipedia), built once and shared by every model. A per-model gate decides which tables each quant actually uses. On Qwen3.8-27B (Lite-Beam 27B) the gated engram lifts AA-Omniscience dev from −22.5 to +4.5 and SimpleQA from −25.5 to +38.5.

Engram 1 (tablinum/beam-engram-tables) shipped with Lite-Beam 1.0 (501B) as three tables (Software, Law, OER). Tablinum-2 adds 19 tables (about 54M passages) of primary sources for finance, health, STEM and humanities, and adds the per-model gate.

Quick links: Tables · How the gate works · Use it · Lite-Beam 27B · License

The idea

Retrieval is model-independent: a static first-stage embedder, BM25 and a small cross-encoder rerank. So the tables are built once. What a table is worth depends on what the model already knows. A table can lift one model and hurt another: Health hurt Beam 501B (6.1 → 3.0) but helps Qwen3.8-27B (−33.3 → +6.1). So each model gets its own route, chosen on a paired dev test, and the recipe is the same for every model.

The gate

For each model and each domain, the same questions are answered closed-book and with the domain's tables. A table is kept only if the paired sign test favours it and incorrect answers do not rise. The final route is checked once on an untouched test split.

Lite-Beam 27B (Qwen3.8-27B), AA-Omniscience dev 200:

domain	tables	closed-book	engram	better / worse	gate
Software	software, wiki	−20.6	+35.3	16 / 2	on
Law	law, wiki	−33.3	+15.2	14 / 2	on
Health	health, healthp, wiki	−33.3	+6.1	13 / 4	on
Finance	papers_fin, finreg, wiki	−24.2	+3.0	11 / 3	on (provisional)
STEM	papers_sci, wiki	−6.1	−6.1	7 / 7	off
Humanities	papers_hum, philo, wiki	−17.6	−26.5	3 / 6	off

A cheaper retrieval-only screen runs first: is the gold answer in the top-3 / top-10 passages at all? It drops tables that add no coverage before any LLM time is spent. Broad reference and Q&A corpora (encyclopedias, Stack Exchange) added almost nothing on AA's narrow questions. Paper abstracts, case law and EU law did.

Tables

Released = passages, FAISS IVF index and BM25 index included, as tables/<name>.tar. Recipe = the exact build script and source list in recipes/; the text is not redistributed (source terms, or health data kept local by policy). Every table, released or not, has a recipe, so the whole engram can be rebuilt for private use.

table	source	passages	license of the text	here
software	DevDocs (842 docs)	4.83M	per-doc upstream licenses	in Engram 1
law	US case law (Caselaw Access Project)	15.35M	CC0 / public domain	in Engram 1
finreg	Congressional Research Service reports	1.24M	CC BY 4.0	released
eulex	EUR-Lex finance / tax / company law	6.57M	© European Union, reuse with attribution	released
ecfr	eCFR titles 12 17 21 26 29 31 42 45	0.23M	US public domain	released
uscode	US Code, all titles (govinfo)	0.37M	US public domain	released
humref	Encyclopaedia Britannica 1911	0.32M	public domain	released
sestem sefin sehum selaw	Stack Exchange science / finance / humanities / law	4.83M	CC BY-SA 4.0	released (share-alike)
health	MedRAG textbooks, PubMed abstracts, StatPearls	27.39M	mixed; not redistributable	recipe
healthp	clinical guidelines (epfl-llm/guidelines)	0.76M	per-source terms	recipe
fdalabel	openFDA prescription labels	0.27M	public domain	recipe (health: local)
cdcj	CDC MMWR / EID / PCD	1.19M	CC0	recipe (health: local)
papers_fin papers_hum papers_sci papers_eng papers_bio	Semantic Scholar abstracts (peS2o v3)	22.19M	ODC-By; abstracts © publishers	recipe
philo	Stanford Encyclopedia of Philosophy	0.44M	© Stanford	recipe
math	nLab, ProofWiki, Stacks Project	0.34M	mixed (nLab unlicensed)	recipe
oer	open textbooks (Common Pile)	17.50M	CC BY / CC0	in Engram 1 (not routed)
wiki	English Wikipedia	57M	CC BY-SA 4.0	recipe (size)

Every passage starts with its source tag and citation, e.g. 26 U.S.C. §1. Tax imposed, EU law 32015R0035, FDA label: Brand (generic) - clinical studies. The model can see where a number comes from, and BM25 can match cited identifiers.

Use it

pip install faiss-cpu tantivy pyarrow tokenizers safetensors transformers torch huggingface_hub
git clone https://huggingface.co/datasets/tablinum/tablinum-2 && cd tablinum-2
mkdir -p tables && tar -xf tables/eulex.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=["eulex"])
for h in eg.retrieve("What is the minimum liquidity coverage ratio under EU rules?", k=3, tables=["eulex"]):
    print(h["table"], h["text"][:200])

code/engram_eval.py runs a closed-book or engram arm against any OpenAI-compatible server (--url, --route v2 --tables auto), and code/grade.py grades it. Build a recipe-only table with recipes/<script>.sh (see recipes/README.md).

License

The Tablinum-2 compilation, meaning the table design, routes, gates, indexes, manifests 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 (column above). The Stack Exchange tables are CC BY-SA 4.0 and must stay share-alike. Upstream attribution is in SOURCES.md. Recipe-only tables carry no text here; building them is subject to each source's terms.

Citation

@misc{litebeamengram2026,
  title  = {Tablinum-2: the Lite-Beam engram},
  author = {Tablinum},
  year   = {2026},
  note   = {Per-model gated retrieval tables for quantized LLMs},
  url    = {https://huggingface.co/datasets/tablinum/tablinum-2}
}