K-Decision-9B v2

Korean-specialized decision model by ONYULTEA

In collaboration with Hallym University

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 β€” all in a single forward pass. System One / Jev compatible.

Benchmarks

Benchmark Score
LocalLLaMA typed-decisions (2,000 decisions, fitted) 0.8295
KoBEST 5-task conversion (2,600 items) 0.866
Korean Life-Domain Holdout (7 domains) 0.997
korean-jev-decision-index (3,000 items) 0.635
Decision Index 0.3 public suite (140K requests) Index 30.7 Β· 0 errors Β· p50 168ms
  • Trained and served on a single 24GB consumer GPU (int8 LoRA, r=32)

Quickstart

/v1/systemone server

git clone https://github.com/onyultea/k-decision
python src/serve_systemone.py --merged ONYULTEA/K-Decision-9B-v2 --port 8111
// Request
{"state": "Civil complaint: Strong gas smell (call center)",
 "questions": {
   "route":  {"type": "choice", "instructions": "Which department?",
              "criteria": {"gas": "Gas Safety", "water": "Waterworks", "road": "Roads"}},
   "urgent": {"type": "noul", "instructions": "Urgent?"}}}
// Response β€” 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}}}

Acknowledgments

We sincerely thank Prof. Youngwoong Ko (School of Software, Hallym University) for his invaluable guidance and support throughout this fine-tuning research. This work was conducted as an industry–academia collaboration between ONYULTEA and Hallym University. In addition, Hallym University regulations were used as reference material for the compliance-decision training and evaluation in the development and research of K-Decision.


K-Decision-9B v2 (ν•œκ΅­μ–΄)

ONYULTEA ν•œκ΅­ νŠΉν™” λ””μ‹œμ „ λͺ¨λΈ

ν•œλ¦ΌλŒ€ν•™κ΅ μ‚°ν•™ν˜‘λ ₯

μƒνƒœ(state)와 νƒ€μ΄ν•‘λ“œ μ§ˆλ¬Έμ„ μž…λ ₯λ°›μ•„ λͺ¨λ“  μ˜΅μ…˜μ— λ³΄μ •λœ ν™•λ₯  뢄포λ₯Ό 단일 forward 패슀둜 λ°˜ν™˜ (System One / Jev ν˜Έν™˜).

μ„±λŠ₯

벀치마크 점수
LocalLLaMA typed-decisions (2,000κ²°μ •) 0.8295
KoBEST 5νƒœμŠ€ν¬ 0.866
ν•œκ΅­ μƒν™œ 도메인 (7도메인) 0.997
korean-jev-decision-index 0.635
Decision Index 0.3 (14만 μš”μ²­) μ§€μˆ˜ 30.7 Β· μ—λŸ¬ 0

λΉ λ₯Έ μ‹œμž‘

git clone https://github.com/onyultea/k-decision
python src/serve_systemone.py --merged ONYULTEA/K-Decision-9B-v2 --port 8111

κ°μ‚¬μ˜ κΈ€

λ³Έ μ—°κ΅¬μ˜ νŒŒμΈνŠœλ‹ κ³Όμ • μ „λ°˜μ— 걸쳐 κ·€ν•œ μ‘°μ–Έκ³Ό 지원을 보내주신 **ν•œλ¦ΌλŒ€ν•™κ΅ μ†Œν”„νŠΈμ›¨μ–΄μœ΅ν•©λŒ€ν•™ κ³ μ˜μ›… κ΅μˆ˜λ‹˜**께 μ§„μ‹¬μœΌλ‘œ κ°μ‚¬λ“œλ¦½λ‹ˆλ‹€. λ³Έ μ„±κ³ΌλŠ” (μ£Ό)μ˜¨μœ¨ν‹°μ™€ ν•œλ¦ΌλŒ€ν•™κ΅μ˜ μ‚°ν•™ν˜‘λ ₯으둜 μ§„ν–‰λ˜μ—ˆμŠ΅λ‹ˆλ‹€. λ˜ν•œ λ³Έ K-Decision의 κ°œλ°œΒ·μ—°κ΅¬μ—λŠ” ν•œλ¦ΌλŒ€ν•™κ΅ κ·œμ •μ΄ κ·œμ • μ€€κ±° νŒλ‹¨(regulation compliance decision) ν•™μŠ΅Β·ν‰κ°€μ˜ μ°Έμ‘° 자료둜 ν™œμš©λ˜μ—ˆμŠ΅λ‹ˆλ‹€.

Β© 2026 ONYULTEA Γ— Hallym University Β· https://onyultea.kr

Downloads last month
73
Safetensors
Model size
9B params
Tensor type
F16
Β·
Inference Providers NEW
This model isn't deployed by any Inference Provider. πŸ™‹ Ask for provider support

Model tree for ONYULTEA/K-Decision-9B-v2

Finetuned
Qwen/Qwen3.5-9B
Finetuned
(1,483)
this model

Space using ONYULTEA/K-Decision-9B-v2 1

Evaluation results

  • accuracy (fitted) on LocalLLaMA/typed-decisions (test, 400 cases / 2,000 decisions)
    self-reported
    0.830