tablinum/beam-engram-tables

Beam Engram Tables

This is a bundle repository: the table data lives in the tar archives. The dataset viewer shows data/table_index.parquet, a compact index of the four tables; download the archives for the actual passages and indexes.

The retrieval layer that added +7.5 AA points to Lite-Beam 1.0. These are the Software, Law, and OER tables plus the code used by the release route for tablinum/Lite-Beam-1.0-501B-A23B-IQS_XXS.

• Model: Lite-Beam-1.0-501B-A23B-IQS_XXS
• Space: lite-beam-1.0-501b
• Methodology: how the route and scores were made

The gain

AA-Omniscience test 400	Base Q2	Base + S+L engram
index	16.25	23.75
correct / incorrect	112 / 47	140 / 45

On the 133 held-out Software+Law items: 18.05 → 40.60, +22.6, p=0.0001. The route keeps only the tables that fit: Software and Law. Health and OER did not add on this prompt and are off the AA path.

Full engram arm

The same retrieval layer now has full max-effort results alongside the short base screens:

benchmark	base Q2 screen	Q2 + engram	note
GPQA Diamond 198	92/198 (46.5%)	173/198 (87.4%)	final; 25/39 retry items correct; 194 valid (3 errors, 1 length)
AIME 2026 30	9/30 (30.0%)	29/30 (96.7%)	final; four old 8K caps recovered, fifth hit the 32K cap
AA-Omniscience 400	16.25	23.75	paired, minimal reasoning
Software+Law 133	18.05	40.60	paired, minimal reasoning
SimpleQA dev 200	6.5	50.5	optional wiki table, not part of this bundle
HLE text-only MC 560	not published	82/560 (28 correct)	running at max / 32K; 5 HTTP errors; current sample, not final

GPQA/AIME base rows are short minimal screens (512 / 1,024-token caps); the engram rows are full max-effort 32K runs. Those two comparisons are mixed-effort, not paired quant-retention measurements. The new scorecard is in the model card and results Space.

Tables

table	passages	IVF	BM25	archive
software	4,827,247	1.3 GB	359 MB	software-engram-table.tar
law	15,354,819	3.9 GB	2.0 GB	law-engram-table.tar
OER	17,500,842	4.4 GB	2.2 GB	oer-engram-table.tar
health	27,390,442	6.8 GB	3.5 GB	build recipe in health/

Each archive contains passages/ (parquet text), index/ (Faiss IVF + OnDisk lists), and bm25/ (Tantivy). The software archive carries docs.json with per-doc attribution; the OER archive carries files.txt.

Health/StatPearls is built locally but its source terms do not permit redistribution of the passages, so the exact build recipe is included in health/ instead. Nothing from the health table is sent to a remote endpoint.

Quick start

Unpack the tables under tables/<name> and use the code in code/:

from engram import Engram

eg = Engram(tables=["software", "law"])
for hit in eg.retrieve("How do I configure a PHP JSON exception?", k=10,
                       tables=["software"]):
    print(hit["score"], hit["title"], hit["text"][:160])

Release route:

• Software Engineering → software + root wiki
• Law → law + root wiki
• everything else → closed-book

The release run used NPROBE=512, NCAND=100, HYBRID=1, k=3, v0 injection, and cross-encoder reranking. The wiki root table is not included; point INDEX at your own root table, or run a single table.

Coverage

Software is 836/842 unique DevDocs documents. The six missing versions are d3~5, d3~6, gtk~3.22, numpy~1.22, qt~5.14, qt~6.1; five URLs were live at build time and gtk~3.22 returned 404. The table is complete for the available DevDocs corpus; the six versions were not re-indexed.

License

Code is Apache-2.0. Table passages retain their source licenses; see SOURCES.md for DevDocs, CAP, and the open-textbook sources. Keep the notices with any redistribution.

Citation

@misc{beamengram2026,
  title  = {Beam Engram Tables: Software, Law, and OER},
  author = {Lightspeed / tablinum},
  year   = {2026},
  note   = {Retrieval tables for Lite-Beam-1.0-501B-A23B-IQS\_XXS}
}