K-Decision-9B v0.1

Korean-specialized decision model by ONYULTEA — first public release.

K-Decision is a lightweight decision engine that takes a state and typed questions as input and returns structured decisions with calibrated probability distributions over all options — all in a single forward pass. Unlike chat LLMs, it does not generate text; it predicts only the answer-symbol logits for each question (System One / Jev compatible).

Benchmarks

Benchmark	Score	Scale
LocalLLaMA typed-decisions (test 400 cases / 2,000 decisions, fitted)	0.818	Agent traces · Security · CS · Invoice (4 domains, 5 questions each)
KoBEST 5-task conversion (boolq/copa/hellaswag/sentineg/wic)	0.866	2,600 items
Korean Life-Domain Holdout	0.997	Admin(20 services) · Medical · Shopping CS · Life Infra · Education · Workplace · Agent (7 domains)
korean-jev-decision-index	0.635	64 Korean benchmark sources, 3,000 items
University Regulation Compliance	0.791	Novel task (3 regulation documents)
Decision Index 0.3 public suite	Index 30.7 (raw 44.8)	140K requests, 0 errors, p50 latency 168ms

• Trained and served on a single 24GB consumer GPU (int8 LoRA, r=32)
• Calibration: ECE 0.001 level (domain holdout), per-type temperature fitting

Quickstart

1) Hugging Face transformers

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model = AutoModelForCausalLM.from_pretrained(
    "ONYULTEA/K-Decision-9B", torch_dtype=torch.bfloat16, device_map="auto")
tok = AutoTokenizer.from_pretrained("ONYULTEA/K-Decision-9B")

2) /v1/systemone server (System One / Jev compatible)

git clone https://github.com/onyultea/k-decision
python src/serve_systemone.py --merged ONYULTEA/K-Decision-9B --port 8111

// Request
{"model": "k-decision",
 "state": "Civil complaint: Strong gas smell reported (call center)",
 "questions": {
   "route":  {"type": "choice", "instructions": "Which department handles this?",
              "criteria": {"gas": "Gas Safety Division", "water": "Waterworks", "road": "Road Maintenance"}},
   "urgent": {"type": "noul", "instructions": "Is this an urgent case?"}}}

// Response — probability for every option
{"answers": {
   "route":  {"type": "choice", "choice": "gas",
              "probabilities": {"gas": 0.87, "water": 0.05, "road": 0.08}},
   "urgent": {"type": "noul", "noul": 0.93}}}

Question types: choice (2–26 options), noul (yes/no). Compatible with Decision Index reproduction kit and Kev clients.

Method

1. Marker single-pass — A JSON skeleton {"route":□,"urgent":□} is appended to the prompt; the logit immediately before each marker predicts that field's answer. All fields decided in one forward pass (no decoding loop).
2. Constrained decoding — Softmax restricted to legal answer symbols (A–Z / Y,N) → guaranteed sum-to-1 probabilities for all options.
3. Calibration — Per-type (choice/noul) temperature fitted on held-out data to optimize Brier/ECE — probabilities are trust scores.
4. int8 LoRA training — 4-bit quantization was measured to degrade probability precision; int8 adopted instead (+3pp accuracy at same budget).

Training Data

• Korean life-domain synthetic (~28K records): 20 government services, medical triage, shopping CS, life infrastructure (real estate/transport/telecom), education, workplace, agent tool gating — with honorific/casual register, typo, and slang augmentation.
• General typed decisions: tasksource-jev-typed-decisions (CC-BY-4.0) sample, multilingual distillation (Apache-2.0, teacher-probability soft labels), tool-decision pool (CC-BY-4.0).
• Regulation compliance task: public university regulation documents (topic routing, numeric-threshold violation, applicability).

Limitations

• Choice up to 26 options, state up to 6,000 characters — exceeding limits returns Unsupported (HTTP 422, honest rejection).
• World knowledge (KMMLU ~30%) is at 9B base level — fine-tuning changes expression, not knowledge.
• High-stakes decisions (medical treatment, credentialing, hiring) require human review.
• High probability does not guarantee correctness — threshold-based operation recommended.

License

• Model weights: Apache-2.0 (derived from Qwen3.5-9B)
• Training & evaluation code: github.com/onyultea/k-decision

Citation

@software{kdecision2026,
  title = {K-Decision-9B: Korean-specialized decision model},
  author = {ONYULTEA},
  year = {2026},
  url = {https://huggingface.co/ONYULTEA/K-Decision-9B}
}

---

K-Decision-9B v0.1 (한국어)

ONYULTEA 한국 특화 디시전 모델 — 첫 공개 릴리스.

K-Decision은 상태(state)와 타이핑드 질문(typed questions)을 입력받아, 모든 옵션에 걸쳐 보정된 확률 분포를 붙인 구조화 결정을 단일 forward 패스로 반환하는 경량 의사결정 엔진입니다. 일반 챗 LLM과 달리 텍스트를 생성하지 않고, 각 질문의 정답 심볼 로짓만 예측합니다 (System One / Jev 호환).

성능

벤치마크	점수	규모
LocalLLaMA typed-decisions (2,000결정, fitted)	0.818	에이전트 트레이스·보안·CS·인보이스 4도메인
KoBEST 5태스크 변환	0.866	2,600문항
한국 생활 도메인 홀드아웃	0.997	행정·의료·쇼핑CS·생활인프라·교육·직장·에이전트 7도메인
korean-jev-decision-index	0.635	64개 한국 벤치마크 소스 3,000문항
대학 규정 준거 판단	0.791	규정 3종 기반 신규 태스크
Decision Index 0.3 공개 스위트	지수 30.7	14만 요청, 에러 0, 지연 p50 168ms

• 학습·서빙 모두 단일 24GB 소비자 GPU (int8 LoRA, r=32)
• 확률 보정: ECE 0.001 수준, 타입별 온도 피팅

빠른 시작

허깅페이스 transformers

from transformers import AutoModelForCausalLM, AutoTokenizer
import torch

model = AutoModelForCausalLM.from_pretrained(
    "ONYULTEA/K-Decision-9B", torch_dtype=torch.bfloat16, device_map="auto")
tok = AutoTokenizer.from_pretrained("ONYULTEA/K-Decision-9B")

/v1/systemone 서버

git clone https://github.com/onyultea/k-decision
python src/serve_systemone.py --merged ONYULTEA/K-Decision-9B --port 8111

질문 타입: choice(2~26옵션), noul(예/아니오). Decision Index 킷·Kev 클라이언트 호환.

작동 원리

1. 마커 단일패스 — JSON 스킬레톤의 마커(□) 직전 로짓이 해당 필드 결정을 예측, N개 필드 1회 forward 동시 결정
2. 제약 디코딩 — 합법 심볼(A~Z / Y,N)만으로 softmax, 합계 1 보장
3. 캘리브레이션 — choice/noul 각각 온도 피팅, Brier/ECE 최적화
4. int8 LoRA — 4비트 양자가 확률 정밀도를 해침을 실측 확인, int8 채택

한계

• choice 최대 26옵션, state 최대 6,000자 — 초과 시 정직 거부
• 세계 지식(KMMLU ~30%)은 9B 베이스 수준 — 파인튜닝은 발현을 바꾸지 지식을 만들지 않음
• 고위험 의사결정에는 인간 검토 필수

© 2026 ONYULTEA · https://onyultea.kr