quyet498/sunny

kyta-decision-v1

Vietnamese decision model (Noul / Choice / Score): unsloth/Qwen3.5-2B + LoRA + decision head, trained with Unsloth FastDecisionModel on quyet498/type-decision-vi.

Training

177,274 training examples (38,902 grouped multi-question examples for batch independence, BAT_AUGMENT = True), LoRA r = 16, learning rate 2e-4, effective batch 32, 1 full epoch(s).

Results

Test split: {'accuracy': 0.8781994704324801, 'ece': 0.010263822440120451, 'loss': 1.1508666963242833, 'record_accuracy': 0.8707667298835005}

JevBench jevbench-mini (0-100):

inference	overall	REP	BAT	MEA	LOG	CHO
batched	70.1	77.5	70.0	44.6	72.1	86.1
isolated (one question per call)	76.6	80.0	99.2	43.1	74.6	86.1

Recommended inference: isolated.

Use

from unsloth import FastDecisionModel
model, tokenizer = FastDecisionModel.from_pretrained("quyet498/sunny", max_seq_length = 2048, load_in_4bit = True)
FastDecisionModel.for_inference(model)

def predict_isolated(state, questions):
    answers = {}
    for key, question in questions.items():
        answers.update(FastDecisionModel.predict(model, tokenizer, state, {key: question}))
    return answers