system-one-gemma
Open Jev-style decision model built on Gemma 3 270M with a scoring head, scoring all options in one forward pass on CPU.
Overview
System One Gemma is an open-source implementation of a System One scoring architecture inspired by Jev, built on Google's Gemma 3 270M because the author had no GPU. For each option it builds a sequence of state, question and option, runs it through the transformer, takes the last token's hidden state through a linear scoring head to get one number, and softmaxes across options, all options batched in one forward pass. It supports Choice, Noul and Score, and a Colab notebook is provided.
The README is candid that accuracy is modest for a 270M model but reports good calibration (ECE 0.047), arguing that a model whose 62% means 62% is useful for thresholds and escalation. It also explains how Jev plays Doom by reading game state as structured text and choosing among actions, as an illustration of the pattern. The README's '200x faster, 200x cheaper' framing is the author's comparison with LLM generation. Code is Apache-2.0, and Gemma's licence must be accepted on Hugging Face.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which support queue? | technical support · billing · general inquiry · sales · cancellation |
| Yes / no | Is this message written in English? | — |
| Score | Rate this review (1-5 stars) | 1 star · 2 stars · 3 stars · 4 stars · 5 stars |
How it works
- Accept the Gemma licence and install requirements
- Build state/question/option sequences
- Score each option with the head in one batched pass
- Softmax to probabilities and threshold
Strengths
- Runs on CPU
- Reported ECE 0.047
- Simple, readable architecture
Caveats
- Modest accuracy at 270M
- Not benchmarked against Jev
Getting started
pip install -r requirements.txt after accepting the Gemma licence on Hugging Face, or open the Colab notebook.At a glance
- Type
- Project
- Status
- research
- Language
- Python
- Runtime
- PyTorch / Transformers
- Has code
- yes
- License
- Apache-2.0
- Stars
- 7
Use cases and integrations
How curated lists describe it
Open-source Jev-style System One decision model. Gemma 3 270M with a scoring head — fast, calibrated decisions in a single forward pass. No text generation. Inspired by TypeSafe.ai's Jev.
wh000wh000/awesome-jev-liveOpen-source Jev-style System One decision model. Gemma 3 270M with a scoring head — fast, calibrated decisions in a single forward pass. No text generation. Inspired by TypeSafe.ai's Jev.
hellogumbo/awesome-jev
Frequently asked questions
How accurate is it?
Modest, per the author, but well calibrated (ECE 0.047).
Does it need a GPU?
No; it was built to run on CPU.
Is it compared with Jev?
Not yet; the README lists that as future work.
Related projects
- TypeAR★ 946
TypeLLM: Jev-inspired type-safe decoding for existing autoregressive LLMs on SGLang, with schema-guaranteed typed outputs.
ProjectbetaPython llama.cpp-based C++ server that turns any LLM into a System One decision model, reading answers from next-token logits.
ProjectbetaC++MLX library that turns an existing local model into a decision engine by reading candidate logits, with no fine-tuning, on Apple Silicon.
ProjectprototypePythonPython library that turns an open LLM into a Jev-style classifier via constrained single-token choice indexes in one forward pass.
ProjectprototypePythonOfficial code for the FastJEV paper: training-free reduction of redundant candidate computation in multimodal JEV-style models.
ProjectresearchPythonLocal server (xks) implementing Jev's /v1/systemone contract on an open-weight LLM, accelerated with Xeon Phi cards.
ProjectresearchRust