karpo-decision-31B / README.md
lionel-mp's picture
Weights licensed under Apache-2.0
29316cf verified
|
Raw History Blame Contribute Delete
2.42 kB
---
title: karpo-decision-31B
emoji: ⚖️
colorFrom: purple
colorTo: blue
sdk: gradio
sdk_version: 6.31.0
app_file: app.py
short_description: Typed decisions from one forward pass — choice, noul, score
python_version: "3.12"
startup_duration_timeout: 1h
---
# karpo-decision-31B — decision model playground
[MachinePulseAI/karpo-decision-31B](https://huggingface.co/MachinePulseAI/karpo-decision-31B) is a 31.3B
decision model, ranked #1 on the [Decision Index](https://huggingface.co/spaces/multimodalart/jev-decision-index)
(Full score 73.9). Fine-tuned from [google/gemma-4-31B-it](https://huggingface.co/google/gemma-4-31B-it),
it replaces text generation with a linear readout: given a **state** — any text or JSON — and a schema of
typed questions, it returns a probability for every option of every question in one forward pass.
## How to use
1. Describe the situation in the **State** box (plain text or a JSON value).
2. Write questions as JSON — one object per question, keyed by question ID. Each question has a `type`
(`choice` / `noul` / `score`), optional `instructions`, and `criteria`:
- **choice** — `criteria` maps option keys to descriptions.
- **noul** — true/false; `criteria` is an optional `{"false": ..., "true": ...}`.
- **score** — `criteria` is a list of at least two ordered level descriptions, lowest first.
3. Press **Decide**. Each answer renders as a probability distribution with the chosen option
(or expected level, or probability of *true*) and a confidence margin; the raw System One JSON
response is in the accordion below.
All questions are answered in a single forward pass — the model reads the state once, and no question
sees another question's options. Text only, up to 32,768 tokens per sequence (a longer request is split
by questions, each part with the full state).
## Notes
- Probabilities are relative to the options you supply: adding or changing an option changes the
distribution.
- The model always picks one of the given options; it never generates text.
- The inference code is the authors' [karpo-decision](https://github.com/machinepulse-ai/karpo-decision)
package (Apache-2.0), vendored in this Space.
Powered by [karpo-decision (MachinePulse)](https://huggingface.co/MachinePulseAI/karpo-decision-31B).
Weights under [Apache-2.0](https://huggingface.co/MachinePulseAI/karpo-decision-31B/blob/main/LICENSE).