systemone-lite
Toy local Jev-shaped decision API on a 0.5B model that answers yes/no, choice and score questions by scoring option tokens.
Overview
systemone-lite is a small local server and Python client that follows the public System One wire shape. Given state and closed questions (noul, choice, score), it returns one discrete answer per question by scoring option tokens on a 0.5B model instead of generating JSON. Weights and the phase-2 training dataset are published on Hugging Face, and the numbers in the README were measured on an RTX 3060.
The author positions it as a trainable experiment for demos and for benchmarking decision heads without a hosted stack, explicitly not affiliated with TypeSafe and not a cloud drop-in. Reported results are modest: 61.6% on the held-out test split and 50.7% on the public JevBench set with ECE 0.245 and about 13 ms median latency, with weak results on game tasks and near-ceiling results on ticket routing. The README warns that calibration is mediocre and confidence should not be treated as a calibrated probability.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Yes / no | Does this need a human agent? | — |
| Choice | Route to a team | billing · technical · other |
| Score | Urgency | low · medium · high |
How it works
- Start the local server with the published 0.5B model.
- POST state and questions to /v1/systemone or use the Python client.
- The model scores option tokens and returns one answer per question.
- Application code acts on the answer, treating confidence cautiously.
Strengths
- Runs on a consumer GPU with short single-question calls.
- Far fewer tokens than generating JSON with the same weights, per the README.
- Open weights and training data.
Caveats
- Calibration is mediocre by the author's own account.
- Weak on game tasks; accuracy well below hosted models.
- Toy project, not a drop-in replacement for Jev.
Getting started
pip install -e ".[dev]" then systemone-lite --model dwidlee/systemone-lite-0.5b --port 8000.At a glance
- Type
- Service
- Status
- prototype
- Language
- Python
- Runtime
- local HTTP server
- Has code
- yes
- Has tests
- yes
- License
- MIT
Use cases and integrations
How curated lists describe it
Toy local System One–style decision API (Jev-shaped). Not affiliated with TypeSafe.
hellogumbo/awesome-jev
Frequently asked questions
How accurate is it?
The README reports 61.6% on its held-out test split and 50.7% on public JevBench, and calls calibration mediocre.
Is it a replacement for Jev?
No, the author describes it as an unofficial local approximation for experiments.
What hardware does it need?
A consumer GPU; reported numbers were measured on an RTX 3060.
Related projects
Small open decision models (0.8B and 4B, from Qwen3.5) with a server that speaks Jev's /v1/systemone format in one forward pass.
ServicebetaPythonServes MiniCPM5-2B-8bit on Apple Silicon as a Jev-compatible /v1/systemone service and reproduces its JevBench score.
ModelprototypePython- reflex★ 166
Small open decision model on a frozen Qwen3.5-4B serving /v1/systemone in about 200 ms, benchmarked against Jev on JevBench.
ServiceresearchPython Open Apache-2.0 Jev-like model family on Qwen3.8-27B that reads one-token candidate preferences and serves a TypeSafe-compatible API.
ServiceresearchPythonOpen MiniCPM5-2B decision model serving /v1/systemone: typed choice, noul and score probabilities in one forward pass.
ModelresearchPythonDocker recipes and ARM64 images for serving ten native decision checkpoints, such as Cloudflare Clef, behind /v1/systemone on DGX Spark.
ServicebetaPython