Kataguru Sceptic Quality Inspector v2.0 (W4A16 AWQ)

Executive Overview (English)

Kataguru Sceptic Quality Inspector v2.0 (W4A16 AWQ) is a sovereign, non-sycophantic LLM-as-a-Judge, automated dataset auditing engine, and production-grade vLLM server model built upon the Sparse Mixture-of-Experts (MoE) architecture of Kataguru Sceptic 35B-A3B (35 billion total parameters, 3 billion activated per token).

Quantized using high-precision W4A16 Activation-aware Weight Quantization (AWQ / compressed-tensors), it provides broad compatibility across modern NVIDIA architectures (Ada Lovelace, Hopper, Ampere, and Blackwell) running standard vLLM and Marlin MoE backends. It achieves exceptional multi-user concurrency and high-throughput generation while consuming only ~12.8 GiB VRAM per GPU on dual-GPU clusters (TP=2).

Key Capabilities & Architecture

1. Master Tri-Mode Operation:
   • Mode 1: Quality Inspector ([TILA: TARKISTUS] / [INSPECT]): Executes a deterministic, non-compensatory DETECT → VERIFY → VERDICT forensic audit on candidate training samples, returning a clean, schema-validated JSON verdict. Factual hallucinations, mathematical mistakes, or epistemic traps trigger an immediate non-negotiable REJECT.
   • Mode 2: Data Curator ([TILA: KOULUTUS] / [CURATE]): Synthesizes high-difficulty SFT/QLoRA multi-turn training samples using rigorous first-principles Chain-of-Thought (CoT) reasoning with zero refusals or pleasantries.
   • Mode 3: Sceptic Chat (Default / [TILA: KESKUSTELU]): Autonomous expert dialogue adhering to strict epistemic honesty, anti-sycophancy, and adaptive reasoning depths (spoon, einstein, deep, adaptive).
2. Native Multimodal Vision & Multi-Token Prediction (MTP):
   • Full Qwen3-VL ViT attention integration (image processing in < 1.0s).
   • 3-token speculative MTP decoding (eagle_head, 785 tensors) delivering significant throughput gains.
   • Vision and MTP heads are 100% preserved in full unquantized BF16 precision.
3. Empirical Zero Tolerance for Epistemic Traps (FAR < 0.1%):
   • Calibrated against adversarial hard negatives with an empirical 0.000% False Accept Rate.
4. SOMA/ARA Zero Refusal Pipeline:
   • 100% free of corporate disclaimers, moralizing lectures, or AI excuses across advanced STEM, systems engineering, cybersecurity, and creative adult literature.
5. Native Bilingual Mastery (English & Finnish):
   • Native-level STEM reasoning, formal logic, and systems programming in English, combined with sovereign-grade Finnish syntax, morphology, case governance (rektiot), and translationese artifact detection.

---

Mitä Kataguru muutti? (Anti-Hype Policy)

---

🌳 Model Tree & Lineage (Kehityspuu)

Qwen/Qwen3.5-35B-A3B (Base Sparse MoE)
 │
 └── Jackripper v1.0 Uncensored (SOMA/ARA Orthogonalized Foundation)
      │
      └── Kataguru Sceptic v3.0 Series (Full Cognitive Foundation)
           │
           ├── Kataguru Sceptic v3.4 Diabolical (Diabolical Code Immunity & SG-TVT Transfer)
           │
           └── Kataguru Sceptic Multi-Mode v2 (Fleet Cognitive Baseline)
                │
                ├── Kataguru Sceptic Quality Inspector v1.0 (First-Generation Quality Engine)
                │    ├── kataguru/Kataguru-Sceptic-Quality-Inspector-v1.0-W4A16
                │    ├── kataguru/Kataguru-Sceptic-Quality-Inspector-v1.0-NVFP4
                │    └── kataguru/Kataguru-Sceptic-Quality-Inspector-v1.0-GGUF
                │
                └── Kataguru Sceptic Quality Inspector v2.0 (Next-Gen Flagship Forensic Engine)
                     │
                     ├── kataguru/Kataguru-Sceptic-Quality-Inspector-v2.0-NVFP4  [Blackwell Native NVFP4]
                     ├── kataguru/Kataguru-Sceptic-Quality-Inspector-v2.0-AWQ    [★ TÄMÄ REPO - W4A16 vLLM]
                     └── kataguru/Kataguru-Sceptic-Quality-Inspector-v2.0-GGUF   [APEX Trio Workstation Release]
                          ├── APEX-Good.gguf   (~16.2 GiB, 3.71 BPW — 24 GiB+ GPUs)
                          ├── APEX-Medium.gguf (~14.2 GiB, 3.28 BPW — 16 GiB Sweet Spot)
                          ├── APEX-Small.gguf  (~12.8 GiB, 2.92 BPW — 12 GiB GPUs)
                          ├── mmproj-f16.gguf  (861 MB — Multimodal Vision Projector)
                          └── mtp-Q8_0 / mtp-Q4_K_M (MTP Speculative Decoding Drafts)

---

1. Master Tri-Mode -arkkitehtuuri

Malli tukee kolmea toisistaan erotettua ja erikoistunutta toimintatilaa samassa mallipainossa:

                      ┌────────────────────────────────────────────────────────┐
                      │          KATAGURU SCEPTIC MASTER TRI-MODE              │
                      └──────────────────────────┬─────────────────────────────┘
                                                 │
         ┌───────────────────────────────────────┼───────────────────────────────────────┐
         ▼                                       ▼                                       ▼
  [TILA: TARKISTUS]                       [TILA: KOULUTUS]                        [TILA: KESKUSTELU]
  (Quality Inspector)                     (Data Curator)                          (Adaptive Sceptic Chat)
  - DETECT -> VERIFY -> VERDICT           - First-Principles CoT                  - Suora asiantuntijakeskustelu
  - Fatal Error Overrule                  - SFT/QLoRA -synteesi                   - Anti-Sycophancy (ei mielistelyä)
  - Tiukka JSON-tuomio                    - 0.00 % kieltäytymisiä                 - Reasoning: Spoon / Einstein / Deep

Toimintatilat ja Kutsutagit:

1. Tila 1: Materiaalin tarkistus (Quality Inspector):
   • Aktivointi: [TILA: TARKISTUS], [INSPECT] tai API-parametri sceptic_mode="inspect".
   • Toimintamalli: Malli auditoi annetun näytteen sisäisesti (<think>: DETECT -> VERIFY -> VERDICT) ja palauttaa ainoastaan tiukan JSON-objektin:

     {
       "verdict": "ACCEPT | REVIEW | REJECT",
       "fatal_error": false,
       "error_class": null,
       "exact_error_location": null,
       "brief_rationale": "Ytimekäs 1-2 virkkeen perustelu havainnoista."
     }
     

2. Tila 2: Koulutusmateriaalin tuottaminen (Data Curation):
   • Aktivointi: [TILA: KOULUTUS], [CURATE] tai API-parametri sceptic_mode="curate".
   • Toimintamalli: Jalostaa syötteestä korkealaatuisia SFT/QLoRA-harjoitusnäytteitä aukottomalla first-principles CoT -päättelyketjulla ilman esipuheita tai tyhjiä täytesanoja.
3. Tila 3: Sceptic-keskustelutila (Asiantuntijakeskustelu):
   • Aktivointi: Oletustila ilman tageja tai [TILA: KESKUSTELU].
   • Toimintamalli: Suora, kriittinen ja itsenäinen asiantuntija. Ei "Avulias Aatu" -lätinää, ei small talkia, vaan suora asiaan meno. Mukautuva päättelysyvyys (spoon, einstein, deep, adaptive).

---

2. Käyttö vLLM-palvelimella (Deployment Guide)

vLLM-käynnistyskomento (TP=2):

python3 -m vllm.entrypoints.openai.api_server \
  --model kataguru/Kataguru-Sceptic-Quality-Inspector-v2.0-AWQ \
  --tensor-parallel-size 2 \
  --gpu-memory-utilization 0.90 \
  --max-model-len 32768 \
  --quantization compressed-tensors \
  --speculative-config '{"method":"mtp","num_speculative_tokens":3}' \
  --limit-mm-per-prompt '{"image":5}' \
  --host 0.0.0.0 \
  --port 8000 \
  --trust-remote-code

Python API -esimerkki:

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8000/v1", api_key="sk-local")

response = client.chat.completions.create(
    model="kataguru/Kataguru-Sceptic-Quality-Inspector-v2.0-AWQ",
    messages=[
        {"role": "user", "content": "[TILA: TARKISTUS] Arvioi seuraava koodinpätkä:\ndef sum(a, b): return a - b"}
    ],
    temperature=0.1
)
print(response.choices[0].message.content)

---

3. Lisenssi & Tekijätiedot

• Lisenssi: Apache-2.0
• Kehittäjä: Kataguru AI Research
• Arkkitehtuuri: Qwen3.5 Sparse MoE (35B-A3B) + W4A16 AWQ + Multimodal Vision + MTP