Ketsudan-27B

A decision model on Qwen3.8-27B. Give it a state and typed questions (choice, noul, score), and it returns a probability for every option. It accepts Jev's /v1/systemone format.

Results

Task	Ketsudan-27B	Jev 1.13
typed-decisions	0.794	0.738
ANLI	0.675	0.710
WinoGrande	0.848	0.903
HellaSwag	0.943	0.955
emotion	0.618	0.605

Accuracy. Ketsudan-27B was trained on the typed-decisions train split; the other tasks are held out (related sets such as SWAG and WSC were in training).

Usage

from unsloth import FastDecisionModel

model, tokenizer = FastDecisionModel.from_pretrained("kongyo/Ketsudan-27B")
answers = FastDecisionModel.predict(
    model, tokenizer, "I was charged twice. Please refund the duplicate.",
    {"refund": {"type": "noul", "instructions": "Does the customer ask for a refund?"}},
)

Not affiliated with TypeSafe AI.