---
title: decider-0.8b
emoji: 🎯
colorFrom: gray
colorTo: blue
sdk: gradio
sdk_version: 6.28.0
app_file: app.py
short_description: Typed decisions with calibrated probabilities, one pass
python_version: "3.12"
startup_duration_timeout: 30m
models:
  - Mapika/decider-0.8b
---

# decider-0.8b

A Gradio demo for [`Mapika/decider-0.8b`](https://huggingface.co/Mapika/decider-0.8b), the smallest member of the decider family (0.75B params on `Qwen/Qwen3.5-0.8B-Base`, bf16, ~1.5 GB).

The model **does not generate text**. It reads a **state** (free text or any JSON value) and a set of **typed questions** — `choice` (2–255 options), `score` (2–10 ordered levels) or `noul` (yes/no) — and returns a **calibrated probability distribution for every question**. The hidden state at each answer slot is projected onto the option-label rows of the LM head and softmaxed over the valid labels.

## Question format

One question per blank-line-separated paragraph. The first line is `<type> <id>: <instructions>`; each `- ` line after it is one option (`name` or `name: description` for choice, one level description for score, lowest first).

```
choice department: Which team should handle this?
- returns: Exchanges, refunds, wrong or damaged items
- billing: Charges, invoices
- other

noul refund_requested: Does the customer request a refund?

score frustration: How frustrated is the customer?
- calm
- frustrated
- very frustrated
```

`Advanced settings` has the softmax temperature (default **1.03**, the calibrated value from the checkpoint's `decider_config.json`), per-question row independence, and isolated Score levels.

## Examples

Taken from the authors' material: the model card's quick start, `examples/routing_with_confidence.py` and `examples/composite_scoring.py` from [github.com/Mapika/decider](https://github.com/Mapika/decider), the demo block in `decider/infer.py`, and a FrozenLake board from the repo's game environments.

## Implementation

`decider/` is the authors' inference subset, copied unchanged from the model repository (Apache-2.0). `app.py` runs the eager branch of `Decider.system_one`: the model is loaded at module scope and moved to CUDA (ZeroGPU), and the temperature is passed in explicitly. The CUDA-graph engine is not used.

Weights and code: Apache-2.0.