trev

trev is a typed decision model: LoRA adapters + an Unsloth Clef decision head on unsloth/Qwen3.5-0.8B.

Input: a state string and typed questions (choice / noul / score).
Output: native option probabilities in one forward pass (no generated tokens).

Files

File	Role
adapter_model.safetensors + adapter_config.json	PEFT LoRA (r=16, α=8)
joint_head.safetensors + joint_head_config.json	Clef decision head
unsloth_decision_config.json	Clef layout + calibrated temperatures
tokenizer* / chat_template.jinja	Tokenizer

Inference-only — no optimizer / trainer state.

Load & predict (Unsloth)

from unsloth import FastDecisionModel

repo = "YOUR_HF_USER/trev"  # or a local path to this folder

model, tokenizer = FastDecisionModel.from_pretrained(
    model_name=repo,
    max_seq_length=32768,
    load_in_4bit=True,
)
FastDecisionModel.for_inference(model)

answers = FastDecisionModel.predict(
    model,
    tokenizer,
    state="Hi, I was charged twice for invoice #4411. Please refund the duplicate today.",
    questions={
        "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?",
        },
    },
)
print(answers)

Requires a recent Unsloth build with FastDecisionModel / Clef support.

Training (summary)

• Base: unsloth/Qwen3.5-0.8B (QLoRA, 4-bit)
• LoRA: r=16, alpha=8, dropout=0.05
• Data: internal ~96k typed-decision training rows
• Export: best validation checkpoint after temperature calibration

License

Apache-2.0 for these adapters and the decision head.
Base model terms follow unsloth/Qwen3.5-0.8B / Qwen.