Is this a trained System One model or a decoding setup?

It is a decoding setup, not a newly trained classifier. The chat template is prefilled with choice_index, sampling is masked to valid indexes, and batching reuses one prefill for multiple question suffixes.

Why was the System One Doom run faster than tool calls?

With the same model and prompt the System One run had much lower median action latency and routinely emitted simultaneous inputs such as strafing while turning and firing, while the tool-call run mostly emitted one input per turn.

How does it handle pages with over a thousand links?

Numeric indexes struggled past a thousand links, so labels worked better there. The Wikipedia demo jumps from Baseball to Sun in three hops using that label path.

When should I enable cache_prefix batching?

Leave it off for one or two questions and enable it above three questions, where prefilling once and evaluating all suffixes together avoids repeating the prefill step per question.

How do I install SystemOne and make a call?

Install with pip install -e ., then build an engine with SystemOne.from_pretrained("Qwen/Qwen2.5-0.5B-Instruct"). Call engine.system_one(state=..., questions={...}) with Choice questions from typesafe_sdk, and read result.choices["team"].choice and result.choices["team"].probabilities.

What does the Doom demo show about tool calling?

Using the same Qwen3-8B model and prompt, the ordinary tool-call version was about 3.5 times slower between actions (600 ms versus 172 ms median) and tended to do one input per turn, whereas the System One version routinely entered many simultaneous inputs such as strafing, turning, and firing at once.

Can I point it at a model other than the example?

Yes. from_pretrained takes a Hugging Face model name, and the README's example uses Qwen2.5-0.5B-Instruct while the Doom and Wikiracing demos use Qwen3-8B. It turns whatever model you load into a System One-style classifier by constraining its choice-index logits.

CommunityJev Practices & Patterns

System One — sgoedecke

Minimal Python recipe that turns any causal LLM into a Jev-style classifier by prefilling choice_index and constraining sampling to legal indexes in one forward pass.

See highlights

Open github.com

Stars: 37
Language: Python
Published by: Community
Author: sgoedecke
Last verified: Oct 2026

sgoedecke/system-one

User repository on GitHub

View on GitHub

Batched single-token choice inference for open language models, compatible with TypeSafe

Forks: 5
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.

This is one of several Jev listings for the same project. The canonical entry is System One — iamaamir.

About this project

Turn any LLM into a System One model like Jev: a fast general classifier that you can supply a set of questions to and get an answer in a single forward pass. The code is vibe-coded but I wrote this README by hand. I reimplemented the Doom and Wikiracing demos from the Jev release post using Qwen3-8B. Here it is playing Freedoom MAP0. We start with two shotgun shells to force a switch to the pistol, and you can see the model choosing to pick up armor and health later on in the level. Download the MP4.

Quoted from the project's README. Read the full README.

Tagscommunity jevbest catalog-import awesomejev

Works forJev for Classification

Used inJev for Software & DevTools

Languages: Python

Highlights

Single-pass choice via chat template prefill plus logit masking over valid choice indexes.
Batched multi-question path prefills once with cache_prefix, recommended above three questions.
Doom Freedoom and Wikipedia-race demos reimplemented with open Qwen models from source.
Label-based options used for thousand-link pages where numeric indexes performed worse.

Watch out

The author calls the code vibe-coded and hand-wrote the README. Demos need Freedoom assets via ViZDoom and Wikipedia excerpts, with licenses noted. Tool-call comparison used the same model and prompt. Copy reviewed against the repository README on 2026-10-04, not from running it.

Questions about this entry

Short answers drawn from the linked project or page.

More Jev resources like this

jev-on-a-laptop

Unofficial laptop study of parallel constrained decoding: prefill once, share the KV cache across fields, and decide all schema fields in one batched forward pass without writing JSON.