GreenLeaf-Decision-9B

What it does

Given a state (any text or JSON) and schema questions (yes/no, multiple-choice, or score levels), the model returns a calibrated distribution over each question's options in one forward pass. Nothing is generated.

Question types:

• noul — yes/no propositions (2 classes)
• choice — multiple choice, up to 52 options (letters A–Z, a–z)
• score — ordinal levels 0–9

Benchmarks (JevBench public splits)

Re-verified 2026-10-18 end-to-end on this exact repo via the JevBench greenleaf_repo adapter (AutoModel load, model.decide(), single pass, 0 failed, 0 unsupported):

Split	Score	Accuracy	Median latency	Mean latency
JevBench hard	75/111	0.676	0.212 s/item	0.439 s/item
JevBench easy	48/48	1.000	0.097 s/item	0.124 s/item
JevBench original	72/72	1.000	0.099 s/item	0.118 s/item

Usage

from transformers import AutoModel

model = AutoModel.from_pretrained("GreenNode/GreenLeaf-Decision-9B", trust_remote_code=True)

result = model.decide(
    state="The insured reported a water leak that ran 4-6 weeks, concealed inside the wall.",
    question={
        "type": "choice",
        "key": "coverage",
        "instructions": "Which coverage decision applies?",
        "options": [
            "deny_vacancy_exclusion",
            "deny_repeated_seepage",
            "pay_subject_to_10000_sublimit",
            "pay_subject_to_15000_sublimit",
            "pay_full_estimate_less_deductible",
        ],
    },
)
print(result["prediction"], result["probs"])

Prompt layout is Clef's sandwich: system + schema fields, then the state, then the schema fields again — spans are read from the second (state-aware) copy.

License

Apache-2.0. Base model: Qwen3.5; head architecture derived from Cloudflare Clef (Apache-2.0).