Datasets:

dancinlab
/

anima-corpus-5lang-unified

license: cc-by-sa-4.0
language:
  - en
  - fr
  - de
  - es
  - ko
pretty_name: anima 5-lang UNIFIED corpus (wiki + SNS + persona, byte-vocab256)
tags:
  - anima
  - persona
  - sns
  - multilingual
  - byte-vocab
  - unified
  - coverage-corpus

anima-corpus-5lang-unified

A UNIFIED 5-language training corpus that combines a clean encyclopedic wiki backbone with the persona × SNS roleplay surface, so a single byte-level model is multilingual across ALL surfaces (wiki + SNS + persona) in en / fr / de / es / ko. This closes the CORPUS-domain gap where 5-lang coverage previously lived ONLY in the wiki/chat corpus and the persona/SNS corpus was Korean-only.

This corpus feeds a FUTURE 5-lang 7B retrain (separate from the current KR-persona 7B). It is byte-vocab V=256 (UTF-8) throughout.

composition (~50% wiki / ~50% persona-SNS)

part	bytes	%	source	license
wiki backbone	5,218,193	50.05	`wikimedia/wikipedia` 20231101 (en/fr/de/es/ko)	CC-BY-SA-4.0 (real, attributable)
persona × SNS	5,207,508	49.95	authored-synthetic (deterministic templates, 20-roster × 16 scenarios × IG/YT)	authored-synthetic, no PII, no scraped data
unified total	10,485,747	100	block-interleaved (byte-weighted round-robin)	mixed (per part above)

unified sha256: ac6ed840319c503b3045ec997015bd396ecacf58681f79f47fe8d1082adcd995
wiki backbone sha256: 497df619b2c952d5dce158bff990157619dbb8066083f6ae3212312a0811819c
persona×SNS sha256: 1e5a062a5fe216a24ffe8230714d6bc9eb760ac9f3e18eec9326858e659866ba
wiki blocks: 12,268 · persona dialogues: 17,755

per-language byte split (unified, wiki + persona)

lang	total bytes	%	wiki	persona
en	2,007,387	19.14	1,048,576	958,811
fr	2,153,045	20.53	1,048,576	1,104,469
de	2,115,603	20.18	1,048,576	1,067,027
es	2,057,230	19.62	1,048,576	1,008,654
ko	2,152,632	20.53	1,048,575	1,104,057

5-way balanced (19.1–20.5% per language). The wiki backbone is exactly 1 MB/lang; the persona slice rounds-robins uniformly across languages, with ko / fr running slightly larger because Hangul / accented characters cost more UTF-8 bytes per glyph. No silent under-coverage — every language carries both surfaces.

dialogue fraction

~49.95% of bytes are multi-turn dialogue (persona × SNS, 3–8 turns each).
~50.05% is encyclopedic wiki prose (non-dialogue backbone).

philosophy (p2/p3/p4/p6 held)

NO injection in training text. Persona is carried by VOICE only. Turn structure is plain <follower>: … / <persona_name>: … continuation — NO [role:, [persona:, [character: tags. grep of those tags over the unified corpus = 0 (verified).
Per-dialogue metadata (lang, persona_id, platform, scenario, n_turns) lives in a SEPARATE .meta.jsonl sidecar, never interleaved into the training text.
NO synthetic assistant-RLHF padding (p6). The persona lines are authored archetype-voice templates, not cooperation/empathy fine-tuning.

honest scope (a_scale_honest_scope · a_lane_akida_gpu_split N/A)

The wiki part is REAL CC-BY-SA wikipedia text (provenance attributable).
The persona × SNS part is machine-authored multilingual templates — a COVERAGE corpus, NOT native-collected text. The en/fr/de/es lines are authored translations/paraphrases of each archetype's voice (e.g. knight = formal/archaic in every language; ice_queen = cold/sharp), not native-speaker corpora. This is a deliberate coverage scaffold so the 5-lang model has a consistent persona surface in every language; native-collected persona data would be a stronger follow-on, not a claim made here.
Why a fresh backbone (not clm-backbone-5lang-sample): that dataset is ko/en/zh/ru/ja (mC4), off-axis for the en/fr/de/es/ko persona set, and its ko C4 slice contains NSFW/spam web text. This corpus rebuilds a clean, on-axis wiki backbone from wikimedia/wikipedia instead (completeness-bar re-design, a_completeness_over_cheap).

determinism / reproduction

# 1. persona × SNS (5MB, seed 20260604 → identical sha on rerun)
python3 serving/persona_sns_corpus_5lang_gen.py --target-mb 5.0 \
    --out serving/corpus/persona_sns_corpus_5lang.txt

# 2. wiki backbone (1MB/lang, wikimedia/wikipedia 20231101 via datasets-server REST)
python3 serving/build_wiki_backbone_5lang.py \
    --out serving/corpus/wiki_backbone_5lang.txt --mb-per-lang 1.0

# 3. merge → unified (~50/50, deterministic block-interleave)
python3 serving/merge_corpus_5lang_unified.py

The persona generator is fully deterministic (fixed seed; re-run reproduces the same sha256 — verified). The wiki backbone is deterministic modulo the pinned upstream dataset revision (date 20231101). The merge is deterministic.

roster (the 20 personas, archetype voice carried across all 5 langs)

J-anime (0–9): school_idol · senpai · knight · sorceress · noir_detective · horror_whisper · childhood_friend · demon_lord · childlike · stoic_mentor. Korean-webtoon (10–19): ice_queen · chaebol_heir · pure_heroine · tsundere_oppa · airhead_friend · charismatic_prez · thug_returnee · cold_heiress · gentle_oppa · fallen_antagonist. Voice SSOT = HEXAD/VOICE/anima-voice/rp_voice_profiles.hexa.

16 scenarios × {Instagram, YouTube}

팬DM칭찬 · 위로 · 일상잡담 · 고민상담 · 셀카리액션 · 댓글답글 · 라이브Q&A · 추천부탁 · 사과 · 축하 · 응원 · 질문답변 · 일상공유 · 팬아트반응 · 밤인사 · 동기부여. Platforms: instagram_dm/comment/live_qna (~~70%) · youtube_comment/ community (~~30%).

files

persona_sns_corpus_5lang_unified.txt — the unified training text (HF only).
persona_sns_corpus_5lang.txt / .meta.jsonl — persona part + metadata sidecar.
wiki_backbone_5lang.txt — the clean wiki backbone.
persona_sns_corpus_5lang_unified.sample.txt — 200-line sample head.
generators: serving/persona_sns_corpus_5lang_gen.py, serving/build_wiki_backbone_5lang.py, serving/merge_corpus_5lang_unified.py.

cross-links

domains: domains/CORPUS.md (registry + coverage matrix) · [[PERSONA]] · [[SNS]]. KOSMOS enrichment analysis: domains/CORPUS-enrichment-analysis.md.
governance: a_hf_registry · a_hf_collections · a_kosmos · a_scale_honest_scope · p2/p3/p4/p6 · a_completeness_over_cheap.