KoJev-v0

Korean typed-decision encoder. A state plus choice / score / noul questions in, calibrated probabilities out, one forward pass. Not a chat model.

Independent of TypeSafe Jev. Wire shape follows POST /v1/systemone.

License

Apache-2.0. Backbone is a fine-tune of skt/A.X-Encoder-base (Apache-2.0, SKT AI Model Lab). See NOTICE.

What's in the box

file	what
model.safetensors + config.json	A.X Encoder backbone after SFT
head.safetensors	span-pooling grouped-softmax head
kojev_config.json	pooling, temperature, markers
tokenizer.json	A.X tokenizer plus [STATE] [Q] [OPT]

Load with the KoJev package (load_checkpoint), not AutoModelForSequenceClassification.

from pathlib import Path
from huggingface_hub import snapshot_download
from kojev.encoder import load_checkpoint
from kojev.schema import Example, Question, QuestionType

root = Path(snapshot_download("NomaDamas/KoJev-v0"))
model, collator, _ = load_checkpoint(root)
example = Example(
    state="오늘 날씨가 맑다.",
    questions=[
        Question(
            type=QuestionType.NOUL,
            instructions="긍정적인 내용이다.",
            options=["아니오", "예"],
            gold=1,
            meta={},
        )
    ],
    source="demo",
    split="dev",
)
print(model.decide(example, collator))

Training (short)

• Backbone skt/A.X-Encoder-base, full finetune (backbone lr 2e-5, head lr 1e-3).
• Gold mix: 12 Korean HF sources, ~104k train states / ~311k questions. KoBEST held out.
• SFT 1 epoch × 3 seeds, then one continue-train epoch from seed 2. RLCD attempted, NO-GO; this file is SFT-only.
• Product run: sft-full-seed2-ep2, diverged=false.

Honest numbers

In-domain gold-val overall 0.764 (majority 0.645). OOD rule-gold 0.416 (majority 0.482). KoBEST zero-shot is near chance. This is a fitted Korean gold encoder, not a general Korean Jev.

split	acc
gold-val	0.764
ood	0.416
kobest-boolq	0.490
kobest-copa	0.475
kobest-wic	0.501
kobest-hellaswag	0.318
kobest-sentineg	0.505

Same 80-example slices vs OpenRouter typesafe/jev-1.13: Jev wins every KoBEST task by a wide margin; KoJev only competes on gold-val.

Citation

@misc{kojev-v0,
  title={KoJev-v0},
  author={NomaDamas},
  year={2026},
  howpublished={\\url{https://huggingface.co/NomaDamas/KoJev-v0}},
}

Also cite A.X Encoder-base.