decider-2b for mlx-vlm

Mapika/decider-2b v11 (revision 533964da), with its weight names changed from model.language_model.* to model.* so the checkpoint loads through the standard mlx-vlm loader. Tensor values and dtypes are unchanged (bf16), and the config, tokenizer and decider_config.json are copied as-is.

decider-2b is a calibrated decision model by Mapika, fine-tuned from Qwen/Qwen3.5-2B-Base. It returns a probability distribution over the options of each typed question in one forward pass. See the original model card for training, evaluation and limitations.

from mlx_vlm import load
from mlx_vlm.models.decider2 import Decider2

model, processor = load("nativ-community/decider-2b")
decider = Decider2(model, processor)
result = decider.predict(
    {"ticket": "Please refund my duplicate charge"},
    {
        "department": {
            "type": "choice",
            "instructions": "Which team should handle this ticket?",
            "criteria": {"billing": None, "technical": None, "sales": None},
        },
        "refund": {"type": "noul", "instructions": "Does the customer request a refund?"},
    },
)
print(result["answers"])

On the same inputs, this checkpoint gives exactly the same predict output as the original.

License: Apache 2.0, inherited from the original model.