1. Architecture & 254 + 20 = 274 Expert Allocation Table
Table with columns: Expert Range, Language / Domain, Morphological & Linguistic Specialization| Expert Range | Language / Domain | Morphological & Linguistic Specialization |
|---|
#0..#253 (254 Frozen Core) | Universal STEM, Coding, Math & Reasoning (en / Universal) | 100 % bit-identical Kataguru-Sceptic-35B-A3B-v3.1 core (GPQA Diamond 70.20 %, LiveCodeBench v6 39.43 %, BFCL V4 83.67 %, PLC/CNC 96.67 %) |
#254, #255 | Finnish (fi — Suomi) | Modern Kotus standard Finnish (15 cases, consonant gradation, Selkokieli /selko, 0 % 1900s Gutenberg archaisms) |
#256, #257 | Estonian (et — Eesti keel) | EKI standard Estonian (14 cases, three-degree quantity alternation kolm väldet) |
#258, #259 | Hungarian (hu — Magyar) | MTA Hungarian (18 cases, vowel harmony, definite/indefinite conjugation) |
#260, #261 | Lithuanian (lt — Lietuvių) | VLKK Lithuanian (7 cases, archaic Indo-European nominal/participle inflection) |
#262, #263 | Latvian (lv — Latviešu) | Literary Latvian (7 cases, 6 declension classes, Baltic palatalization) |
#264, #265 | Icelandic (is — Íslenska) | Árnastofnun Icelandic (4 cases × 3 genders, U-umlaut / I-umlaut, quirky subject case) |
#266, #267 | Basque (eu — Euskara Batua) | Euskaltzaindia Unified Basque (ergative-absolutive alignment, polypersonal Nor-Nori-Nork auxiliaries) |
#268, #269 | Georgian (ka — ქართული) | Literary Georgian (Kartvelian split-ergativity, polypersonal verb screeves, Mkhedruli script) |
#270, #271 | Welsh (cy — Cymraeg) | Standard Welsh (initial consonant mutations treigladau, VSO syntax, inflected prepositions) |
#272, #273 | Turkish (tr — Türkçe) | TDK Turkish (2/4-way vowel harmony, consonant mutation, deep agglutinative suffixation) |
| Format & Footprint | Universal vLLM W4A16-AWQ (compressed-tensors, group_size=128, 274 experts/layer, ~20.5 GiB on disk, ~19.1 GiB in vLLM --language-model-only — fits on a single 24 GB GPU!) | |
- Direct
L1 -> L2 Translation Mode: Prefix your prompt with /translate /<target_code> (e.g. /translate /et, /translate /fi, /translate /hu, /translate /lt, /translate /lv, /translate /is, /translate /eu, /translate /ka, /translate /cy, /translate /tr, /translate /en, /translate /selko) followed by source text in any supported language.
- Native Cross-Lingual RAG & Grounded Citation Engine (
Mode 7: /rag or documents=[...]):
- Pass
documents=[{"id": "1", "title": "...", "lang": "et", "text": "..."}, ...] to tokenizer.apply_chat_template(...) (vLLM / Hugging Face / LlamaIndex / LangChain standard), OR prefix your prompt with /rag /<target_code> and include <documents>...</documents> or <context>...</context>.
- Reads retrieved chunks in any combination of the 11 languages (
FI, ET, HU, LT, LV, IS, EU, KA, CY, TR, EN) and synthesizes a strictly grounded response in the target language with mandatory inline source citations ([1], [2]), automatic conflict reporting ([1] vs. [2]), and explicit abstention when a requested parameter is absent from the retrieved documents.
- Native Language Expert Chat Mode: Prefix your prompt with
/fi, /et, /hu, /lt, /lv, /is, /eu, /ka, /cy, /tr, /en, or /selko.
# Cross-Lingual RAG Example (Estonian + Hungarian + English Docs -> Grounded Finnish Answer with [1], [2], [3] Citations)documents = [ {"id": "1", "title": "DC-Bus Protection", "lang": "et", "text": "Sagedusmuunduri alalisvoolu vahelüli maksimaalne lubatud pinge on 820 V DC."}, {"id": "2", "title": "SIL3 STO Relay", "lang": "hu", "text": "A SIL3 biztonsági relé 4 ms-on belül feszültségmentes állapotba kapcsolja a szervohajtást."}, {"id": "3", "title": "Spindle Cooling", "lang": "en", "text": "The liquid-cooled CNC spindle requires a minimum glycol-water flow rate of 6.5 L/min at 22 °C."},]messages = [ {"role": "user", "content": "/fi Mitkä ovat välipiirin maksimijännite, STO-turvareleen kytkentäaika ja karan minimivirtaus?"}]prompt = tokenizer.apply_chat_template(messages, documents=documents, tokenize=False, add_generation_prompt=True)
3. Native Sceptic 7-in-1 Multi-Mode, Multimodal Vision & Agentic Architecture
Beyond 110-pair direct translation and cross-lingual RAG, Kataguru-Sceptic-35B-A3B-Polyglot-274E-v3.2 inherits 100 % of Kataguru Sceptic's native multi-mode operating system, 27-layer ViT multimodal vision encoder, <tool_call> AST engine, and graduate STEM/coding core inside a single model (0 server restarts):
Table with columns: Built-In Engine, Trigger Tag / Argument, Capability & Production Role| Built-In Engine | Trigger Tag / Argument | Capability & Production Role |
|---|
| Mode 1: Forensic Quality Inspector | [INSPECT] or sceptic_mode="inspect" | Verdict-first JSON dataset, math, and code auditor ({"verdict": "ACCEPT|REJECT|REVIEW", "fatal_error": false, "learning_value": 9.2, ...}) with <100 ms early-exit rejection. |
| Mode 2: Data Curator & Code Surgeon | [CURATE] / [SURGEON] | Synthesizes high-learning-value (>= 8.5) SFT samples ([CURATE]) or self-heals buggy/truncated code and flawed derivations ([SURGEON]). |
| Mode 3: 20 ms System 1 Reflex Judge | [DECISION] or sceptic_mode="decision" | Zero-output-token (max_tokens=1, logprobs=True) 0–9 / A/B epistemic & quality reflex scorer (>1M samples/day). |
Mode 4: 110-Pair Direct Translator | /translate /<target> | Direct L1 <-> L2 literary & technical translation across FI, ET, HU, LT, LV, IS, EU, KA, CY, TR, EN without an English pivot (0.00 % refusal). |
Mode 5: Multimodal Vision & <tool_call> Agent | Default Chat + images / tools=[...] | 27-Layer ViT Vision (mmproj included), BFCL V4 83.67 % <tool_call> AST, Systems SWE (LiveCodeBench v6: 39.43 %), PhD STEM (GPQA Diamond: 70.20 %), and Industrial PLC/CNC (96.67 %). |
| Mode 6: Empirical Cognitive Modes | {REASON:ihyper|hyper|logic|007|spoon|einstein|off} | Structured MECE (ihyper: +8.69 pp accuracy, -31 % latency) and fault-tolerant agentic state-machine execution (007). |
| Mode 7: Cross-Lingual Citation RAG | /rag /<target> or documents=[...] | Synthesizes retrieved chunks across any mix of the 11 languages into grounded target-language answers with mandatory [1], [2] citations and abstention on missing facts. |
4. Copyright, Provenance & Apache-2.0 Attribution Notice
- Developer & Model Family: Kataguru (
https://huggingface.co/kataguru)
- Copyright & Post-Training IP Holder: Datasein Oy (Finnish Business ID / Y-tunnus: 3167094-7, Finland)
- Chief Architect & Lead Developer: Jari Katajisto, Finland
- Quantized Release License (
Apache-2.0, Section 4(c)–(d) & Section 6): These quantized weights (NVFP4, W4A16-AWQ, GGUF, built upon the open Qwen3.5 foundation) may be used commercially and redistributed under Apache-2.0, provided all Datasein Oy / Kataguru copyright notices and the NOTICE file are retained intact. No trademark rights are granted. Full BF16 master weights, curated training databases (TekijäL 49 § / 96/9/EC), and domain-specialized expert cartridges remain proprietary trade secrets of Datasein Oy (All Rights Reserved).
Suomenkielinen Yhteenveto / Finnish Summary
🏢 Virallinen yritysjulkaisu — Datasein Oy (Y-tunnus: 3167094-7, Suomi)
Hugging Face -profiili @kataguru toimii suomalaisen Datasein Oy:n (Y-tunnus: 3167094-7) virallisena yritystilinä — ei yksityishenkilön tai harrastelijakehittäjän tilinä. Pääarkkitehti Jari Katajisto toimii Datasein Oy:n nimissä, ja kaikki Kataguru-malliston jatkokoulutuksen, kielimoduulien ja kuratoitujen tietokantojen tekijänoikeudet kuuluvat Datasein Oy:lle.
Kataguru-Sceptic-35B-A3B-Polyglot-274E-v3.2 on Datasein Oy:n (Y-tunnus: 3167094-7, pääarkkitehti Jari Katajisto, Suomi) kehittämä Kataguru Sceptic 35B-A3B -arkkitehtuurin 254 + 20 = 274 asiantuntijan monikielinen 7-in-1-monitoimimalli, jossa 254 jäädytettyä STEM-, koodaus-, PLC/CNC- ja päättely-ydineksperttiä (#0..#253) on yhdistetty kymmenen morfologisesti vaativan kielen (fi, et, hu, lt, lv, is, eu, ka, cy, tr) dedikoituihin kielieksperttipareihin (10 × 2 = 20 eksperttiä #254..#273). Malli kääntää suoraan kahden kielen välillä ilman englannin välikieltä (11 × 10 = 110 kieliparia), toimii monikielisenä RAG- ja lähdeviittausmoottorina (/rag tai documents=[...]) pakollisilla [1], [2]-viitteillä sekä sisältää natiivina koko Sceptic-monitoimiympäristön: 27-kerroksisen multimodaalisen vision, <tool_call>-työkaluagentin (BFCL V4: 83,67 %), laaduntarkastajan ([INSPECT]), aineistokuraattorin/koodikirurgin ([CURATE] / [SURGEON]), <20 ms Systeemi 1 -pikatuomarin ([DECISION]) ja {REASON:ihyper} / {REASON:007} -päättelytilat. Julkiset kvantit jaetaan Apache-2.0-lisenssillä (alkuperä- ja NOTICE-merkinnät säilytettävä kohdan 4(c)–(d) mukaisesti); täydet BF16-mestaripainot ja jatkokoulutetut erikoisasiantuntijat ovat Datasein Oy:n yksityistä omaisuutta (All Rights Reserved).