decisions-tiny

A small fine-tuned decision model inspired by the System 1 / typed-decision approach used by models such as JEV. It performs structured decisions from natural-language inputs, including categorical classification and yes/no-style decisions.

Model Details

• Base model: Qwen3.5-0.8B
• Fine-tuning: LoRA with 4-bit base-model loading

Dataset

• Dataset: LocalLLaMA/typed-decisions

Evaluation

Metric	Score
Accuracy	0.7035
Record accuracy	0.2375
Loss	1.0787

Inference

from unsloth import FastDecisionModel

model, tokenizer = FastDecisionModel.from_pretrained(
    model_name="Archit-01/decisions-tiny",
    max_seq_length=2048,
    load_in_4bit=True,
)

FastDecisionModel.for_inference(model)

text = (
    "Hi, I was charged twice for invoice #4411. "
    "Please refund the duplicate today."
)

schema = {
    "team": {
        "type": "choice",
        "instructions": "Which team should handle this?",
        "criteria": {
            "billing": "invoices, payments, refunds",
            "technical": "bugs, outages, errors",
            "sales": "pricing, new plans",
        },
    },
    "refund": {
        "type": "noul",
        "instructions": "Does the customer ask for a refund?",
    },
}

answers = FastDecisionModel.predict(
    model, tokenizer, text, schema
)

print(answers["team"]["answer"])
print(answers["team"]["probabilities"])
print(answers["refund"]["answer"])