alea-institute/alea-synth

ALEA Synth

alea-synth is a synthetic English pretraining corpus from the ALEA Institute. Each document was written by one call to OpenAI gpt-6-luna from a sampled specification and one anchor. Anchors are dictionary senses from OpenGloss, a SHELF train-split bibliographic spec, or a passage of public-domain U.S. legal text from the KL3M corpus. Later anchors were reached by snowball sampling over the concepts that accepted documents actually used. Every document passed deterministic filters, a strict blind LLM judge and decontamination against ALEA's evaluations. Each one carries full per-document provenance and lineage.

• 124,742 documents, 99,324,653 ALEA tokens (tokenizer alea-p1sw-staged-16k) in 4 document configs, plus a lineage config with 5,226,878 snowball frontier edges.
• Built for pretraining and mid-training of small models. ALEA's R1 model trained on most of run1; see R1 usage.
• Generator: gpt-6-luna (service tier flex, reasoning effort low) for both generation and judging.
• License: CC-BY-4.0; see License and attribution for the seed sources.

How to load

from datasets import load_dataset

ds = load_dataset("alea-institute/alea-synth", "kl3m-grounded")            # train + validation
print(ds["train"][0]["text"])

# all document configs as one training stream
from datasets import concatenate_datasets
cfgs = ['opengloss-concept', 'shelf-form', 'kl3m-grounded', 'gaps']
train = concatenate_datasets([load_dataset("alea-institute/alea-synth", c, split="train") for c in cfgs])

# drop the flagged pre-fix "Mara" documents, or keep only what R1 never saw
train = train.filter(lambda r: "name_collapse_prefix" not in r["flags"])
unseen = train.filter(lambda r: r["used_in_r1"]["c2"] == 0 and r["used_in_r1"]["d"] == 0)

edges = load_dataset("alea-institute/alea-synth", "lineage", split="train")   # snowball frontier edges

Configs and splits

config	train	validation	docs	ALEA tokens	token share	tokens per doc
opengloss-concept	51,981	536	52,517	46,095,075	46.4%	230–2,587 (median 869)
shelf-form	20,511	192	20,703	14,103,745	14.2%	250–2,799 (median 540)
kl3m-grounded	26,283	269	26,552	19,390,118	19.5%	179–2,104 (median 696)
gaps	24,699	271	24,970	19,735,715	19.9%	11–3,305 (median 776)
total	123,474	1,268	124,742	99,324,653	100%

• opengloss-concept (run1 family og_concept): 1–3 OpenGloss senses written into one of 21 general forms.
• shelf-form (run1 shelf_spec): a SHELF train-split bibliographic spec (LC class, topics, LCGFT form, register, audience, length, geography), with OpenGloss senses fitted to its topics.
• kl3m-grounded (run1 kl3m_ground): an OpenGloss legal sense plus a KL3M public-domain passage that uses it, rewritten into a teaching or practice form (explainer, worked scenario, client letter, Q&A, memo, case brief, …). Every number and citation must appear in the seed. - gaps: the second, targeted batch (below).
• lineage: one row per snowball edge (used, related, broader, narrower, contrast, lcc_neighbor) from frontier_edges.jsonl. from_item_in_release says whether the source document survived the release fixes.

Splits. Each config has a train split and a small validation split, a documented random ~1% of documents. A document is in validation iff int(sha256("alea-synth/validation/v1/" + id)[:12], 16) < 0.01 · 2^48. The rule is stable across versions, so a document never moves between splits. There is no evaluation split taken from SHELF: SHELF's validation and test splits are held out for ALEA evaluation and were never used as seeds (decontamination rule h below). Note: R1 trained on validation documents too (used_in_r1), so validation is a held-out sample for monitoring new models, not a clean test set for R1.

Columns

column	type	meaning
id	string	24-hex document id (blake2b of spec id + text), unique across the release
text	string	the document: title line, blank line, body (Markdown where the layout calls for it)
family	string	config name (opengloss-concept, shelf-form, kl3m-grounded, gaps)
generator_family	string	the generator's internal family or stream (og_concept, shelf_spec, kl3m_ground, gaps streams)
form	string	document form (112 in run1, e.g. FAQ page, case brief, sermon, model statute)
layout	string	plain_prose, structured_markdown, qa_pairs, transcript_style, verse, argument_map, …
register	string	professional, conversational, casual, plain, formal, academic, journalistic, technical, creative
audience	string	target reader (45 in run1, e.g. small-business owners, ESL learners, paralegals)
length_bucket	string	brief 150–300, short 300–550, medium 550–900, long 900–1,400 words (requested)
anchors	list	anchor senses: sense_id, lexeme_id, headword, pos, gloss (OpenGloss v2.4), domain, tier, role
lcc, lcc_subject, topic	string	Library of Congress class letter, ALEA's own subject description, SHELF/OpenGloss topic
legal_source	string	KL3M source for grounded docs (usc, ecfr, fr, plaw, cap, agreements, reg-docs, fcc, triples), else null
seed	struct	KL3M grounded docs only: source, repo, revision, file, row_group, identifier, char_start, char_end, window_sha256, passage_id, license
judge	struct	judged, verdict, faithfulness 0–3, quality 0–3, flags, version
lineage	struct	wave, parents (ids), edge_type that reached the anchor, edges, root_seeds (og:<sense>, shelf:<doc>, kl3m:<passage>)
alea_tokens	int	length in ALEA alea-p1sw-staged-16k tokens (no specials)
used_in_r1	struct	passes over this doc in ALEA R1's stage mixes: c2, d
flags	list	name_collapse_prefix, clock_time_opening, scrub_placeholder, generated_during_run_pause (see Release fixes and Known limitations)
provenance	struct	model_requested, model_served, service_tier, reasoning_effort, response_id, created_utc, prompt_template_id, system_variant, instructions_sha256, input_sha256, code_git_sha, sampler_version, judge_version, spec_id, craft_slots, shelf_doc_id, generation_attempts, repairs, run

How it was made

Seeds (wave 0)

Wave-0 anchors come only from ALEA's own and public-domain data:

• OpenGloss v2.4 senses (mjbommar/opengloss-v2.4-*). The primary sense is stratified by domain root (15) × part of speech (noun .5 / verb .2 / adj .2 / adv .1) × frequency band (core / tier 2–3 / tier 4–5). Up to 2 co-anchors are taken along its own relations.
• SHELF train-split specs (mjbommar/SHELF, config all, revision fba779ad). Non-text forms such as maps, films and recordings are skipped. Authoritative legal forms (statutes, contracts, opinions) are written as teaching models with invented names.
• KL3M passages, used when the anchor is a legal sense. A passage pool was range-read, a few row groups per repo, from the pinned revisions below. Windowing: USC/eCFR sections cut at the editorial notes, FR windows after SUMMARY / SUPPLEMENTARY INFORMATION, PLAW windows at SEC. boundaries.
  • reg-docs and FCC windows must look like a real public comment, and they were scrubbed of personal data before prompting: signature blocks, "I, ," lines, e-mail, phone, street/PO box, city+ZIP and links.
  • Each KL3M passage grounds at most one document.

Spec sampler

Each document gets one spec, drawn independently of the anchor except where a rule ties them:

• Form. 21 concept forms (feature, essay, story, handout, lesson plan, podcast, dialogue, letter, diary, op-ed, exhibit text, field notes, speech, news report, FAQ, case study, review, poem, interview, how-to, encyclopedia entry). SHELF's own LCGFT forms come with SHELF specs. For KL3M items a grounded operation is chosen by source (explainer, worked scenario, client letter, Q&A, memo, interview, FAQ, training brief, summary + outline, argument map (comments only), debate, handout, newsletter column, case brief (CAP), plain-language contract summary (EDGAR)).
• Register. 8 registers, plus creative for stories, poems and diaries only.
• Audience. From a legal pool for LC classes K/J/H, otherwise a general pool. A low-grade audience (target Flesch-Kincaid ≤ 10) never gets an academic, technical or formal register, and the prompt states the target grade.
• Length. brief / short / medium / long with weights .12 / .28 / .35 / .25. Verse is always brief; Q&A forms are short or medium.
• Craft slots. 1–3 writing-craft slots from the vendored book-template writing ontology (hooks, closers, paragraph shapes, discourse moves, figures, sentence constructions, diction controls, argument arrangements, evidence devices), filtered so that no "fault" slot is ever drawn.
  • Slot-fit rule. Grounded rewrites draw only form-level slots. Arrangements and evidence devices invite new content, so only concept documents may draw them.
• Labels. LC classes reach the prompt only as ALEA's own subject descriptions, never as LCC captions or SHELF text.
• System prompts. Two frozen system-prompt variants (provenance.system_variant).

Snowball algorithm

After a document passes every filter:

1. Concept extraction. A deterministic longest-match lookup of OpenGloss headwords and inflections runs over the body. It keeps the anchors plus the top 8 other concepts. Ranking: tier specificity × multiword bonus × (1 + ln count) × same-domain bonus. Senses are chosen by Lesk overlap plus a domain prior. Stopwords, proper nouns, retired lexemes and ALEA's held-out evaluation lexemes are skipped.
2. Relations. From each used concept, up to 2 targets per edge type are followed over live OpenGloss relations: synonym/see-also → related, hypernym/holonym/instance-of → broader, hyponym/meronym → narrower, antonym/confusable/contrast paragraphs → contrast.
3. LC neighbours. The document's LC class adds data-driven neighbour classes (lcc_neighbor): cosine similarity of class topic profiles in SHELF train > 0.3, or the top 2.
4. Coverage control. The next wave picks an edge type uniformly, then a concept weighted by 1/sqrt(frontier in-degree) × 1/log2(2 + graph degree) (hub down-weighting) × 3 if never used before.
   • Each sense anchors a limited number of documents.
   • Legal concepts with an unused KL3M passage become grounded items, and lcc_neighbor edges become SHELF-spec items in the neighbouring class.
   • The rest become concept documents, which keep the concept they were reached from as a co-anchor with probability .45.
5. Lineage. Every document records wave, parents, the edges and root_seeds, so it traces back to wave 0. All edges are in the lineage config.

Run1 used 5 waves (0–4). Documents per wave in this release: w0 15,997, w1 17,446, w2 16,071, w3 34,765, w4 40,463. In run1, 52,837 distinct anchor senses were used in opengloss-concept alone (66,211 across run1 at generation time), with a top-50 anchor share of 2.2% and a Gini of 0.46. The share of anchors new to each wave was 100% / 92% / 72% / 62% / 50%.

Filters (fail-closed) and the strict judge with repair

Every verdict is stored per item in the generation logs. In order:

1. parse. A Title: envelope and a completed response.
2. length. 0.5–1.6× the bucket midpoint.
3. leakage. Assistant tells, references to "the source/passage/material", spec-label lines, SHELF self-labels, craft-slot name echoes, and dictionary-style label echoes of anchors.
4. glyph. No LaTeX and ≤ 0.5% other non-ASCII. Argument-map glyphs only in argument_map.
5. fidelity. Grounded docs: every U.S.C./C.F.R./§ cite, Pub. L. number, dollar amount, date and multi-digit number must appear in the seed, with one corrective retry. Concept docs: no statutory citations at all.
6. concept. Every anchor lexeme occurs, and Lesk does not clearly prefer another sense.
7. copy. ≤ 60% of the body's word 13-grams may be shared with the seed.
8. readability. Flesch-Kincaid grade ≤ the audience maximum + 1.5.
9. pii. E-mail, phone, SSN, street address and PO box are rejected, as are names recorded from the seed.
10. judge. A strict, blind gpt-6-luna judge (judge-v2-strict).
    • It scores faithfulness (or sense accuracy) and quality 0–3.
    • It sets six explicit fail flags: source_mention, invented_legal_facts, wrong_sense, spec_noncompliant, assistant_tone, other_defect. Any flag or a score < 2 fails the document.
    • The judge was calibrated on the pilot. From 22:23 UTC (kl3m/shelf) and 23:44 UTC (all families) every deterministic pass was judged.
    • A failed document gets one repair pass guided by the judge's evidence, then a fresh judgment. Readability-only failures also get one plain-language repair before judging. Repairs are listed in provenance.repairs.
11. decontam. See Decontamination.
12. dedup. Exact (normalized text), MinHash near-duplicate (word 5-grams, 128 permutations, LSH 16×8, Jaccard ≥ 0.8) and spec_id.

Run1 acceptance: 100,668 of 122,281 generated items accepted (82.3%). By family: og_concept 84.0%, shelf_spec 85.9%, kl3m_ground 77.0%. The judge ran 96,877 judgments and passed 92.9%. Rejections by first failing filter:

filter	rejected	share of items
judge	5,874	4.8%
concept	5,682	4.6%
leakage	5,386	4.4%
readability	2,064	1.7%
fidelity	909	0.7%
decontam	554	0.5%
pii	415	0.3%
parse	299	0.2%
length	292	0.2%
glyph	137	0.1%
copy	1	0.0%

The most frequent judge flags were spec_noncompliant (1,952), invented_legal_facts (1,747), other_defect (1,539), wrong_sense (1,087), source_mention (762) and assistant_tone (483).

Pause rules. A cell (family × form, source or audience) whose acceptance fell below 60%, or whose judged failure rate exceeded 8%, was paused automatically until fixed. Three cells stayed paused at the end: shelf_spec|sacred-text commentary, shelf_spec|set of riddles and src:reg-docs.

Cost. Run1 cost $64.05 in API spend, including calibration (266,468 calls, 0 HTTP 429s): about $0.00063 per accepted document.

QA process

Run1 had 23 QA checkpoints, at every wave end and every 5,000 accepted documents. Each one reported acceptance by cell, judge flags, concentration, distribution drift, invented-name concentration and a personal-data scan, followed by a 15-document personal read. The full log is in this repo as QA_LOG.md; three real third-party names and an office address that it quoted from seeds are redacted. Issues found and fixed during the run:

• Weak early judge and acceptance (cp001). Several cells (src:cap, src:fr, kl3m_ground|explainer, some audiences) fell below 60% acceptance. Fix: the judge-guided repair pass and the readability repair; later all families were judge-gated.
• Source-reference tells ("The excerpt does not …") that a pilot judge passed. Fix: a contextual regex and the strict judge-v2 with an explicit source_mention flag.
• Invented-name collapse (cp008). "Mara" appeared in 14.5% of accepted documents, plus Maya/Eli/Venn/Lena/Jonah. Fix (00:09 UTC Oct 2): each brief carries seeded given-name and surname suggestions from a multi-origin pool and bans the collapsed names. After the fix Mara/Maya are at 0.0%. The pre-fix documents are kept and flagged here.
• FR contact officials (cp016). An FR training brief repeated an agency contact's name, office room and street from the Federal Register notice. A scan found seed person-like names repeated in 31% of FR-grounded output. Fix (01:47 UTC Oct 2): every person-like name in an FR window is banned from the output, and the prompt asks for roles. Earlier FR documents were handled at slice and release time (below).
• Clock-time openings (cp017). About 1.5% of documents open with "At 2:13 a.m." and similar; "2:13" appears in 90 documents. Below threshold, not acted on; flagged here.
• Invented place names (cp022). After the person-name fix, "Bellwether" (2.2%) and "North Quay" (0.7%) recur. Not acted on.
• Weakest pairing. SHELF spec × OpenGloss anchor pairs sometimes force anchors into an unrelated piece (for example "gallows/gibbet" in a middle-school lecture on a border).

Release fixes (applied for this publication)

fix	documents	action
FR doc repeats a person-like name from its own seed window (the run1 rule: 452 slice1 + 283 already excluded from slice2)	735	dropped
any other doc containing a name on the run1 FR-name block list (819 leaked seed names minus a reviewed allowlist of statute, place, company and historical names)	5	dropped
release-wide personal-data scan (e-mail, phone incl. en-dash forms, SSN-like, street address, PO box)	138	dropped
person-like name within ~250 characters of an address, room/mail-stop, phone or e-mail marker (mostly FR/eCFR "send comments to , Room …" passages)	18	dropped
exact duplicates (normalized text)	0	dropped
near duplicates (MinHash, verified word-5-gram Jaccard ≥ 0.8)	0	dropped
pre-fix "Mara" documents (generated before 00:09 UTC Oct 2)	4,625	kept, flags: ["name_collapse_prefix"]
clock-time opening ("At/By hh:mm" in the first line of the body)	1,101	kept, flags: ["clock_time_opening"]
generated (created_utc) inside a run-level "RUN" auto-pause window that did not stop submissions (see Known limitations): run1	0	kept, flags: ["generated_during_run_pause"]
same, gaps1 (windows 13:30:54–13:32:39 UTC; 13:44:55–13:56:13 UTC, Oct 2)	3,654	kept, flags: ["generated_during_run_pause"]
text contains a seed-scrub placeholder copied from the seed ([address removed], [email protected], …), almost all agency addresses	65	kept, flags: ["scrub_placeholder"]
gaps: excluded by the gaps1 generator's own final pass (final_exclusions.json: short_form_function_word_anchor 12, pii 65, contact_person_with_address 9, seed_person_name 3, run1_fr_block_list 4)	93	dropped
gaps: personal-data scan / contact rule / FR block list	0 / 0 / 0	dropped
gaps: exact / near duplicates (against run1 and within gaps)	0 / 0	dropped

Run1: 100,668 accepted → 99,772 released (896 dropped). A final re-scan of every released document finds 0 personal-data pattern hits and 0 contact-rule hits (gaps: 0 and 0). The patterns are regexes, so they catch contact details, not every name. See Known limitations.

gaps (second batch, gaps1)

gaps1 is a targeted second batch for what run1 lacked. It uses the same generator (gpt-6-luna flex, low effort), the KL3M pool and pinned revisions, the OpenGloss index, SHELF train specs, decontamination, dedup, pause rules and QA checkpoints. Its snowball frontier starts from run1's final frontier (wave 0) and then expands from gaps1's own accepted documents (wave 1). There are four streams (generator_family):

stream	docs	what it is
rule_app	7,503	a USC/eCFR/FR rule applied to concrete invented facts (thresholds, exceptions, deadlines, amounts, percentages, periods); every computation is re-executed in code (calc.py) and every number must come from the rule or a verified check
argument	6,201	point/counterpoint, majority vs dissent, steelman, fallacy spotting, valid argument forms, decision memos, weighted pros/cons (weighted totals verified in code)
conversation	6,268	messy multi-party talk: meetings, panels, depositions, hearings, intake calls, negotiations, customer service, e-mail threads with quoted replies, forum threads, chat logs, raw ASR-style transcripts
short_form	4,998	standalone items under 150 ALEA tokens (headlines, definitions, captions, posts, texts, notices, docket entries, citation strings, abstracts, form labels, …), generated 5 per call, each its own document

• Shape. Some gaps documents have no title line, and some carry footnotes.
• Filters. Run1's filters with per-stream adjustments: structure checks for conversations, a token cap for short forms, and calc verification for rule application.
• Judge. The strict judge adds a reasoning_error flag (judge-gaps@…); short forms are judged five per call.
• Size. 24,970 documents and 19,735,715 ALEA tokens after the release fixes below, from 25,063 accepted by the generator.
• R1. All gaps documents are linked 3× in R1's stage-D (decay) mix (used_in_r1 = {"c2": 0, "d": 3}).
• Decontamination. gaps1 batch passes: 52 rule hits (h_shelf_valtest 38, e_likelihood 9, a2_headword 2, d_heldout_novel 2, f_release_eval 1).

Decontamination

Every document was checked before acceptance in per-batch passes, and again as whole training slices, against:

• the ALEA evaluation suite: E1 held-out lexemes and headword forms (rule a2), the knowledge probe (b), ICL suites v1/v2 (c), P1 held-out novel documents by 13-gram (d), the likelihood suite (e), and the OpenLemma eval release (f), using decontam.py from ALEA's OpenLemma pipeline;
• SHELF validation and test (rule h): any shared word 13-gram with SHELF all @ fba779ad val/test (25,104 docs).

E1 held-out lexemes and knowledge-probe words were also kept out of anchors and of the snowball frontier. Batch passes rejected 554 run1 items (598 rule hits: h_shelf_valtest 379, e_likelihood 154, a2_headword 53, f_release_eval 11, b_knowledge_probe 1; an item can hit several rules). The whole-slice passes before R1 tokenization dropped 0 more. Rule (i), an eCFR held-out split, was not implemented because that split does not exist yet.

Statistics

Forms (top 30 of 151)

form	docs	share
explainer	4,807	3.9%
FAQ page	4,792	3.8%
classroom handout	4,417	3.5%
radio interview	3,733	3.0%
speech	2,848	2.3%
poem	2,802	2.2%
podcast transcript	2,784	2.2%
diary entries	2,760	2.2%
how-to guide	2,715	2.2%
question-and-answer set	2,674	2.1%
op-ed	2,655	2.1%
case study	2,594	2.1%
magazine feature	2,585	2.1%
book review	2,575	2.1%
encyclopedia-style entry	2,573	2.1%
short story	2,558	2.1%
field notes	2,531	2.0%
letter to a friend	2,522	2.0%
personal essay	2,514	2.0%
dialogue between two characters	2,485	2.0%
worked example	2,456	2.0%
museum exhibit text	2,394	1.9%
lesson plan	2,179	1.7%
news report	2,097	1.7%
office memorandum	1,972	1.6%
scenario analysis	1,860	1.5%
staff training brief	1,617	1.3%
irac	1,500	1.2%
summary and outline	1,494	1.2%
client advisory letter	1,464	1.2%

Registers

register	docs	share
conversational	23,817	19.1%
professional	20,910	16.8%
plain	15,468	12.4%
academic	13,015	10.4%
casual	12,493	10.0%
journalistic	12,415	10.0%
formal	12,031	9.6%
technical	11,901	9.5%
creative	2,692	2.2%

Audiences (top 25 of 53)

audience	docs	share
general readers	11,266	9.0%
small-business owners	7,757	6.2%
journalists	7,706	6.2%
the general public	6,376	5.1%
retirees	5,414	4.3%
compliance officers	5,387	4.3%
law students	5,215	4.2%
paralegals	5,160	4.1%
high-school students	5,136	4.1%
agency staff	4,897	3.9%
new associates at a law firm	4,772	3.8%
ESL learners	4,705	3.8%
city council members	4,697	3.8%
engineers	3,544	2.8%
nurses	3,050	2.4%
undergraduates	3,038	2.4%
parents	3,004	2.4%
teachers	2,961	2.4%
middle-school students	2,732	2.2%
software developers	2,722	2.2%
hobbyists	2,701	2.2%
farmers	2,632	2.1%
tenants	2,201	1.8%
scientists	765	0.6%
children	699	0.6%

Requested length

bucket	docs	share
medium	41,167	33.0%
short	35,851	28.7%
long	21,843	17.5%
brief	20,883	16.7%
short_form	4,998	4.0%

LC class

LCC	docs	share
K	34,970	28.0%
H	17,591	14.1%
J	12,606	10.1%
Q	9,654	7.7%
G	7,335	5.9%
P	6,091	4.9%
T	5,778	4.6%
L	4,951	4.0%
R	4,648	3.7%
S	3,964	3.2%
B	3,190	2.6%
Z	2,955	2.4%
N	2,416	1.9%
M	2,365	1.9%
C	2,017	1.6%
V	1,142	0.9%
D	1,041	0.8%
U	1,019	0.8%
E	398	0.3%
F	386	0.3%
A	225	0.2%

KL3M seed sources (kl3m-grounded)

source	dataset	docs
ecfr	eCFR (kl3m-data-ecfr)	4,219
agreements	SEC EDGAR agreements (kl3m-data-edgar-agreements)	4,191
usc	U.S. Code (kl3m-data-usc)	3,883
cap	Caselaw Access Project opinions (kl3m-data-cap)	3,860
triples	USC section + knowledge-graph triples (usc-knowledge-graph-triples)	3,118
fr	Federal Register (kl3m-data-fr)	2,350
plaw	Public Laws (kl3m-data-govinfo-plaw)	2,342
fcc	FCC ECFS comments (fcc-submissions)	1,915
reg-docs	regulations.gov public comments (kl3m-data-reg-docs)	674

ALEA tokens per document

tokens	docs
0–199	5,006
200–299	1,359
300–399	17,234
400–499	7,393
500–599	16,904
600–799	20,582
800–999	22,082
1,000–1,199	13,119
1,200–1,499	12,313
1,500–1,999	7,696
2,000–2,999	1,050
≥ 3,000	4

R1 usage

ALEA R1 (ALEA's first model run) used run1 through two tokenized slices of accepted.jsonl, in the alea-p1sw-staged-16k uint16 format:

• slice1: every document accepted by 00:38 UTC Oct 2. 40,763 docs, 32.2M tokens.
• slice2: every later document, minus 283 FR documents that repeated a seed person's name. 59,622 docs, 48.1M tokens.

stage	what was linked	passes
C2	slice1, all 40,763 docs (including the 452 FR-name and 4,629 "Mara" docs)	2
D (decay)	slice2	2
D (decay)	slice1 minus the 452 FR-name docs, keeping a seeded 25% of the 4,629 "Mara" docs (36,819 docs)	1 (slice1's third)
D (decay)	gaps1, all 24,970 released gaps docs (experiments/r1/d-data/add_gaps.py)	3

used_in_r1 = {"c2": passes, "d": passes} per document. The values are the passes in each stage's data mix (mix_manifest.json). At release time R1 was still training (C2 under way, D chained after it), so these are the planned passes; a stage that is stopped early sees fewer. The D slice1 selection was re-derived here and matches the D build's part sha256 (ed7460f4…). Counts in this release:

used_in_r1	docs
c2=0,d=2	59,521
c2=0,d=3	24,970
c2=2,d=0	3,492
c2=2,d=1	36,759

That is 63,425,682 token-passes in C2 and 183,514,265 in D for the released documents. Every released run1 document is in at least one stage; {"c2": 0, "d": 0} would mark documents R1 never trained on; there are none in this release. The dropped FR documents include 452 that R1 saw twice in C2.

Known limitations

• Legal share is 27%, not the 35% target. Run1's family mix was og 52.2% / shelf 20.6% / kl3m 27.3% of accepted documents. Supply of legal-primary anchors with unused passages limited kl3m_ground. Released token shares are in the config table.
• Length range. Run1 documents are 179–2,799 ALEA tokens (94.1% between 300 and 1,999; mean 798). Nothing is longer than ~2.8K tokens, so this is not a long-context corpus.
• Clock-time openings. About 1.5% of documents open with a clock time ("At 2:13 a.m. …"). They carry the clock_time_opening flag.
• SHELF topics are lightly used. SHELF specs supply form, class and audience, but the topic is mostly carried by fitted OpenGloss anchors. Topic × anchor pairing is the weakest part (see QA).
• Name collapse. 4,625 documents generated before the 00:09 UTC fix contain the invented name "Mara" (14.5% of documents at the time; Maya/Eli/Venn/Lena/Jonah were also over-used). They are fine individually and are kept with flags: ["name_collapse_prefix"]. Filter or down-weight them as R1's D stage did (25% kept). Later documents over-use a few invented places ("Bellwether", "North Quay").
• FR names (left over in run1, now fixed).
  • Before the 01:47 UTC fix, FR-grounded documents could repeat Federal Register contact officials and other named individuals. Slice2 excluded 283 of them at tokenization; slice1 had 452, which R1 saw in C2.
  • This release drops all 735, every other document that contains a block-list name, and every document that pairs a person-like name with an address, room or phone marker.
  • Names in the public record that are integral to the source are kept: CAP case parties and witnesses, PLAW honorees, public officials named in their official capacity, and officers or counsel named in public EDGAR agreements (notice clauses with a full address were dropped). The regex and heuristic scans cannot guarantee the absence of every real name.
• Invented legal facts. Grounded documents passed the fidelity filter and the judge, but the judge still flagged invented_legal_facts in 1.8% of judged run1 items. Treat the texts as training text, not legal advice or a legal source. Concept documents in law use invented parties.
• Synthetic-style artefacts. The texts have the regularities of one model family: titles on every document, tidy structure and some template tics. Mix them with natural text.
• Run-level pause bug. r1synth's pause logic (scale.Run.is_paused) never checked the run-wide "RUN" key that the QA rules set on a top-50 anchor share > 10% or a personal-data finding, so such a pause did not stop submissions. The gaps1 run found and fixed it (cp003). Run1 used the same code, but its logs show no RUN auto-pause at all (the top-50 anchor share peaked at 8.7%, below the 10% rule, and all 23 personal-data scans found 0), and operator stops (pause_run) did work, so 0 run1 documents are affected. In gaps1 the RUN key fired twice (top-50 anchor share 11.0% and 11.8%); the 3,654 gaps documents created inside those windows passed every filter and the judge and are kept with flags: ["generated_during_run_pause"]; they mainly over-represent a few frequent legal anchor senses.
• English and U.S. legal context only. No EU OJ or other non-U.S. legal text was used as a seed in run1.

Generator disclosure

• Generation and judging: OpenAI gpt-6-luna through the Responses API, service_tier="flex", reasoning effort low, max_output_tokens 6,000. provenance.model_requested / model_served, service_tier, response_id and created_utc record each document's final generation call.
• Judge: the same model, blind (it is not told the document is synthetic). The judge version is in judge.version and provenance.judge_version.
• Code: prompts and code are pinned by provenance.instructions_sha256, input_sha256, prompt_template_id, code_git_sha (ALEA's alea-model-paper repo, experiments/r1-synth) and sampler_version.
• Clearance: ALEA's project lead cleared gpt-6-luna outputs for training ALEA models. Users should review the generator's terms for their own use.

Intended uses

• Pretraining and mid-training (annealing/decay) data for small language models, especially ones that need coverage of word senses, varied registers and forms, and plain-language explanations of U.S. legal text.
• Research on synthetic-data generation: the provenance and lineage support ablations by family, form, wave, judge score or snowball edge type.

Not intended as a source of legal information, as an evaluation benchmark (use SHELF val/test or ALEA's eval suites), or for training models to impersonate real people.

License and attribution

The dataset is released under CC-BY-4.0. The seeds are credited as follows:

• OpenGloss v2.4 (mjbommar/opengloss-v2.4-*, CC-BY-4.0, Michael J. Bommarito II): senses, glosses, relations, contrasts, encyclopedia text. The anchors.gloss field quotes OpenGloss glosses.
• SHELF (mjbommar/SHELF, CC-BY-4.0, Michael J. Bommarito II): train-split bibliographic specs and LC-class topic profiles. No SHELF text was given to the generator.
• KL3M Data (alea-institute/kl3m-data-*): U.S. federal government works, which are in the public domain in the U.S. under 17 U.S.C. §105. Sources: USC, eCFR, Federal Register, Public Laws, and the triples, which are derived from USC (usc-knowledge-graph-triples).
  • CAP: published court opinions (public domain).
  • EDGAR agreements: public SEC filings.
  • Public comments and submissions: regulations.gov docket comments (kl3m-data-reg-docs) and FCC filings (alea-institute/fcc-submissions @ 22a345f6). Both were submitted through federal public-comment systems whose terms tell submitters that what they submit becomes part of the public record and is made publicly available. ALEA treats these submissions as public-domain material. As an extra precaution, the seed windows were scrubbed of personal data before prompting, and the outputs refer to commenters by role. No commenter names or contact details are reproduced.
• EU Official Journal: not used in run1.
• book-template writing-craft ontology (MIT): only the slot names and definitions were used, to steer the prompts. Ontology git sha and file hashes are recorded per document in the generation logs.

KL3M repo	revision
alea-institute/kl3m-data-usc	97120a0d
alea-institute/kl3m-data-ecfr	e254bec2
alea-institute/kl3m-data-fr	f4929499
alea-institute/kl3m-data-govinfo-plaw	5b41d4fa
alea-institute/kl3m-data-cap	93003ac2
alea-institute/kl3m-data-edgar-agreements	0d62e003
alea-institute/kl3m-data-reg-docs	d51cb245
alea-institute/fcc-submissions	22a345f6
alea-institute/usc-knowledge-graph-triples	8f25e6ca

Files

• data/<config>/{train,validation}-*.parquet: documents (zstd, shards ≤ ~250 MB).
• data/lineage/*.parquet: snowball frontier edges (edge_id, run, wave, type, from_item, from_item_in_release, from_sense, to_sense, og_relation, anchor, score, from_class, to_class, sim).
• QA_LOG.md: the run1 QA log (23 checkpoints and the final summary; a few seed-quoted names redacted).
• QA_LOG_gaps1.md: the gaps1 QA log.
• stats.json: every number in this card, plus the release-fix report.

Citation

@misc{alea_synth_2026,
  title        = {{ALEA Synth}: A Snowball-Sampled Synthetic Corpus Seeded from {OpenGloss}, {SHELF} and {KL3M}},
  author       = {Bommarito, Michael J., II and {ALEA Institute}},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/alea-institute/alea-synth}},
  note         = {Generated with OpenAI gpt-6-luna; CC-BY-4.0}
}

Please also cite OpenGloss, SHELF and the KL3M Data Project when you use the corresponding configs.