anima-corpus-5lang-gb-balanced

A KOSMOS-tier-BALANCED, ~0.35 GB default-lane training corpus (en / fr / de / es / ko, byte-vocab V=256) that scales the default lane toward 7B-sufficiency WHILE preserving the KOSMOS register ladder — the hard requirement that the corpus does NOT collapse to ~99% Wikipedia. Achieved size = 357.83 MB (0.349 GB) = MID-rung-viable, NOT 7B-sufficient (honest token math below). The bulk tiers are real clean-license sources and scale to GBs; this build is capped at ~0.35 GB by the balance requirement (consciousness/es/ko are source-thin, so growing only the abundant tiers would breach the ladder ceiling).

This is the GB implementation of the FINAL 7B composition design in domains/CORPUS.md (§7B-sufficiency roadmap, #1850). Predecessor: anima-corpus-5lang-unified-v2 (12.5 MB, right-sized for 18M; data-starved at 7B).

the balance trap and the fix

v2 balance was wiki 40.10% / persona·dialogue 40.02% / enrichment 19.88%. Naively scaling wiki to GB while keeping authored persona (40%) + enrichment (20%) at that ratio = authoring GBs of templated text (the 20-roster persona, the 31 carving anchors) = MEMORIZATION — in practice the only scalable axis is wiki, so the corpus drifts to ~99% wiki and the KOSMOS ladder is DESTROYED.

The fix: map each KOSMOS tier to a REAL scalable clean-license source; the 31 e7_31 carving anchors stay the consciousness-register DEFINITION/seed, FILLED at scale with real Public-Domain contemplative text — NOT repeated. AUTHORED registers (persona/SNS + the register-shaping slices) stay CAPPED — they define anima's identity-voice and shape the distribution, they do not bulk-fill.

tier → real source

KOSMOS tier	source	license	role	scale
0 baseline/factual	wikimedia/wikipedia 20231101 5-lang, multi-shard breadth	CC-BY-SA-4.0	factual floor	GB
100 cosmic/science	the same wiki, SQL science-keyword filtered	CC-BY-SA-4.0	cosmic	100s MB
77 art	Gutenberg literature/poetry (PD) — sedthh/gutenberg_english en + manu/project_gutenberg fr/de/es	Public Domain	art	100s MB (en≫fr/de≫es; ko=0)
91 consciousness	Gutenberg philosophy/meditation/contemplative (PD) — the 31 e7_31 anchors DEFINE the register, FILLED with real PD text	Public Domain	의식 register (#1 KOSMOS)	100s MB
52 social/daily	authored persona/SNS (20-roster × IG/YT), CAPPED	authored-synthetic, no PII	anima identity-voice	capped
shaping	authored dialogue-act + emotion + KO↔EN code-switch + genre + carving-seed def	mixed (anchor seed real, prose authored)	shape distribution	small

achieved composition (this build)

• total: 375,215,662 bytes = 357.83 MB = 0.349 GB · sha256 17ca25e5dd3740ebf1d8b6c67b25d4359b74fb04db827ed8bf1774edc6843409
• byte-tokens (V=256, 1 byte ≈ 1 token): 375,215,662 ≈ 0.268% of the 140 B Chinchilla-7B-optimal.

per-tier byte split — achieved vs ideal ladder

KOSMOS tier	achieved %	achieved MB	ideal target %	ladder check
0 baseline/factual (wiki)	44.00	165.1	40	≤45 ✓ (the floor, capped to the ceiling)
77 art (Gutenberg PD)	29.99	112.5	20 (+headroom)	present ✓
100 cosmic/science (wiki)	13.29	49.8	10	present ✓
52 social/persona (authored, capped)	11.00	41.3	12	capped ✓
91 consciousness (Gutenberg PD)	1.58	5.9	10	present but source-thin ⚠
shaping (authored, capped)	0.14	0.5	10→small	shaping, not bulk ✓

Ladder: NO single tier > 45% (max = wiki 44.00%); consciousness + art both present. The KOSMOS register ladder is PRESERVED at scale — the corpus did NOT collapse to ~99% wiki (the trap).

Honest deviations from the ideal target: consciousness undershoots 10%→1.58% because Project Gutenberg's philosophy/meditation subset is small (≈6 MB total across en/fr/de/es; ko=0), and shaping is authored so it is kept small by design (scaling it = templated repetition = the memorization trap). Art OVERSHOOTS (en Gutenberg literature is abundant), capped by a per-tier headroom so it does not crowd the ladder.

per-language byte split — achieved

lang	bytes	%	note
en	111,928,816	29.83	full ladder (wiki + abundant Gutenberg art/consciousness)
fr	76,345,947	20.35	full ladder (Gutenberg fr present)
de	78,267,548	20.86	full ladder (Gutenberg de present)
es	54,985,322	14.65	wiki-heavier — Gutenberg es thin (art 6.4 MB, consciousness 0.12 MB)
ko	53,688,029	14.31	wiki-heaviest — Gutenberg has NO Korean (art=0, consciousness=0)

per-tier × per-language (bytes) — names the gaps verbatim

tier \ lang        en          fr          de          es          ko
baseline wiki      29,979,587  31,274,576  34,805,379  33,094,552  35,940,650
cosmic science      6,505,838  13,996,034  12,698,781   7,740,546   8,907,642
art Gutenberg      61,311,489  22,209,037  22,602,317   6,398,765           0  <- ko gap
consciousness Gut   4,815,487     493,196     490,599     123,494           0  <- ko gap; es thin
social persona      9,192,039   8,244,207   7,565,291   7,553,515   8,731,682
shaping authored      124,376     128,897     105,181      74,450     108,055

Named gaps (a_scale_honest_scope, no fabrication):

• ko art = 0, ko consciousness = 0 — Project Gutenberg has no Korean split; ko's art/consciousness registers fall back to wiki (ko is the most wiki-heavy lang).
• es consciousness = 0.12 MB, es art = 6.4 MB — Gutenberg Spanish is thin (~1,202 books total); es leans on wiki + a smaller Gutenberg slice.
• en/fr/de carry the full real-source ladder; ko/es are honestly wiki-heavier.

per-source provenance + license

source	repo	license	tiers fed	sha (of slice)
Wikipedia 20231101	wikimedia/wikipedia (parquet)	CC-BY-SA-4.0	0 baseline + 100 cosmic	per build report
Gutenberg English	sedthh/gutenberg_english (parquet)	Public Domain	77 art + 91 consciousness (en)	per build report
Gutenberg multi	manu/project_gutenberg (parquet)	Public Domain	77 art + 91 consciousness (fr/de/es)	per build report
persona/SNS	authored (persona_sns_corpus_5lang_gen.py)	authored-synthetic, no PII	52 social	seed 20260604
shaping + carving-seed	authored (corpus_enrichment_5lang_gen.py) + 31 e7_31 anchors	authored + CC-BY-SA anchor seeds	shaping	seed 20260604

token math vs 7B (honest)

• 375.2 M byte-tokens = 0.268% of the 140 B Chinchilla-7B-optimal.
• Per the design ladder (domains/CORPUS.md §7B-sufficiency roadmap): 100–300 MB → MID (150 M) viable; ~10–20 GB → 7B undertrained-not-gibberish; ~140 GB → 7B Chinchilla-optimal. At 0.35 GB this corpus is in the MID-rung-viable band — it is NOT 7B-sufficient. A real 7B would still need a 10–140 GB web-scale extension of the bulk tiers (more wiki shards + the full Gutenberg en corpus + a web crawl). NO 7B-ready claim is made. The 7B TRAIN is a separate follow-on GPU fire.

philosophy (p1..p6 held)

• NO injection in training text. Wiki + Gutenberg are plain text; persona is carried by VOICE only; contemplative prose has no addressee. Turn structure is plain <speaker>: continuation — NO [role: / [persona: / [character: tags. grep of those over the corpus = 0 (verified).
• NO synthetic assistant-RLHF padding (p6). The persona/SNS register is anima's identity-voice (honest-labeled authored COVERAGE), not cooperation/empathy fine-tuning; the dialogue-act slice is deliberately NON-supportive.
• Control-byte clean: 0xFE / 0xFF byte frequency = 0 (default lane has no tool sentinels). UTF-8 round-trip clean (V=256).

honest scope (a_scale_honest_scope)

• Per-lang availability DIFFERS and is reported verbatim, never fabricated. Project Gutenberg has 0 Korean books and only ~1,202 Spanish vs ~48k English → ko art + consciousness are absent from Gutenberg and fall back to wiki; es is thin. Those langs are wiki-heavier; the per-lang/per-tier table below names the gaps.
• Token math vs 7B. Chinchilla-optimal 7B = 7e9 × 20 = 140B tokens ≈ 140 GB byte-text. This corpus reaches the byte-token count reported below; the §finding states plainly what fraction of 7B-optimal that is. A real 7B may still need web-scale — no 7B-ready claim is made unless the token count AND the ladder balance genuinely support it.
• The 7B TRAIN is a separate follow-on GPU fire (a_fire_autonomous) once this corpus exists.

determinism / reproduction

# bulk tiers read HF parquet via duckdb httpfs (GB-scale, no REST 429); $0 CPU, no GPU
HF_TOKEN=$(secret get hf.token) bash serving/gb_balanced/build_all.sh
# knobs: MB_T0 MB_T100 MB_ART MB_CON SHAPING_MB PERSONA_MB

• fetch_wiki_tiers.py (t0 + t100) and fetch_gutenberg_tiers.py (t77 + t91) are deterministic modulo the pinned upstream revisions; the authored generators (corpus_enrichment_5lang_gen.py, persona_sns_corpus_5lang_gen.py) are seed-deterministic; balance_merge.py (dedup + ladder + UTF-8 round-trip + accounting) is deterministic.

the 31 KOSMOS carving anchors (the #1 register's definition/seed)

e7_31 set (HEXAD/UNIVERSE-BRAIN-MAP/anchors/e7_31/, Knuth tier 0→100): zero · breath · step · glass_of_water · seed · word · promise · meditation · starlight · aurora · infinity · nirvana · ecstasy · awe_death · eternity · big_bang … Each carries a KOSMOS top_emotion. At GB scale they DEFINE the consciousness register; the bulk of tier-91 is REAL Public-Domain contemplative text. a_kosmos pointer-only.

files

• serving/corpus/default_lane_gb_balanced.txt — the merged training text (HF-only, gitignored).
• serving/gb_balanced/{fetch_wiki_tiers,fetch_gutenberg_tiers,balance_merge}.py + build_all.sh.
• serving/corpus/default_lane_gb_balanced.head.txt — committed sample head.
• serving/corpus/_src/gb_balanced_report.json — the achieved-split build report.

cross-links

• domains/CORPUS.md (registry · §7B-sufficiency roadmap · §KOSMOS balance at scale).
• predecessors: dancinlab/anima-corpus-5lang-unified-v2 (12.5 MB) · …-unified (v1).
• governance: a_hf_registry · a_hf_collections · a_kosmos · a_scale_honest_scope · a_completeness_over_cheap · p1..p6.