User repository on GitHub
Creating system one model just like jev. with different experimentation
- Forks
- 0
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.
CommunityJev Practices & Patterns
Open System One
An Apache-2.0 attempt at an open Jev: a System-1 decision model returning calibrated probabilities per option, with option branches packed so answers do not change when the option list is reordered.
- Stars
- 2
- Language
- Python
- Published by
- Community
- Author
- Readyaddy
- Last verified
- Oct 2026
About this project
Open System-1 is an Apache-2.0 research project building an open alternative to TypeSafe AI's Jev. A System-1 model does not write answers. It receives a situation (the context), a question, and a set of options, and returns a calibrated probability for each option, fast enough to sit inside software as a decision layer. It supports the three decision types Jev defines: Trained on 2026-09-27/28 on a single Colab A100 for 2.4 hours. Full report: experiments/exp9scale/RESULTS.md.
Quoted from the project's README. Read the full README.
Works forJev for Classification
Used inJev for Software & DevTools
- Languages
- Python
Highlights
- Option branches share one prefix but never see each other, so 0.0% of answers change when options are shuffled, down from 12–21% in earlier experiments.
- Held-out scores are reported next to Laya and Jev, including where it loses: 48.6 on the hard JevBench tier against Jev's 74.1.
- A frozen Qwen3-Embedding-4B backbone with LoRA; 4.92B total parameters, 0.9B trained, scored with a proper scoring rule.
- A negative-results section: injected option embeddings score at chance, and a second pass through the decision head changes nothing.
Watch out
The trained checkpoint is not in the repository — the trained half is 1.8 GB in bf16, above GitHub's file limits, and a Hugging Face release is only described as planned, so there is nothing to download yet. Reproducing means rebuilding roughly 331k examples from public sources and training on a single A100-class GPU; the reported run took a couple of hours. Treat the benchmark table as indicative: this model's JevBench numbers come from the public half of each tier while the Laya and Jev columns come from a leaderboard run over the full tiers. Typed decisions are uneven, with choice accuracy well below yes/no. 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.
The held-out table is honest about where it loses: it trails Jev on Banking77, AG News and the hard JevBench tier, and trails Laya on AG News. The comparison is still not apples to apples, and the README says so. Its JevBench scores come from the public half of each tier under the official protocol, while the Laya and Jev columns come from the JevBench v1.2 leaderboard over the full tiers. Read it as a direction of travel, not as a ranking you can act on.
The branch packing. Each option is rendered as its own branch that can attend to the shared prefix and to itself but never to another option, with all branches sharing position ids, so an option's score cannot depend on the set or the order of the others. A test in the repo checks this against scoring each option alone and agrees to a relative difference around 2.3e-6. The measured effect is that zero percent of answers change when options are shuffled, down from 12 to 21 percent earlier in the project.
You have to train it. The checkpoint is deliberately absent from the repository: the trained portion is 1.8 GB in bf16, over GitHub's per-file limit, and the frozen Qwen3-Embedding-4B backbone is fetched from Hugging Face separately. A Hub release is described as planned, with no artifact to point at. The training corpora are also not stored — a prepare script rebuilds about 331k examples from their public sources — so the full path is a multi-hour GPU run rather than an install.
On the numbers in this repository, no — it is behind Jev on two of six held-out benchmarks and ahead only on one. The case for it is licensing and access: Apache-2.0, with the architecture, the data build and the training objective all readable. The stronger engineering claim is behavioural rather than competitive — branch packing makes the model order-invariant, and the objective is a proper scoring rule, so the reported confidence is meant to be usable for escalation decisions rather than decorative.
exp9a: a frozen Qwen3-Embedding-4B backbone adapted with rank-64 LoRA, a decision head with branch packing, and an RLCD-style proper-scoring objective. It is 4.92B parameters total, 0.90B trained, from one Colab A100 run of 2.4 hours.
About 331k examples rebuilt from public datasets across intent, multiple choice, yes/no, ordinal and label-diversity families. The corpora are not stored, roughly 1.3 GB, so prepare_data.py rebuilds and audits them from their sources.
Injecting a precomputed option embedding scored at chance with the vector alone and no better with text, the recurrent rethinking passes changed nothing after the first, and ProofWriter multi-step training taught a surface cue instead of depth.
More Jev resources like this
38GitHub stars
Local unofficial /v1/systemone adapter that serves Choice, Score, and Noul probabilities via first-token LLM judgments or a fine-tuned ModernBERT scorer, including Japanese docs.
Jev Practices & Patterns#community#awesomejev#catalog-import
Community
Argos1111/jev_localPython
Argos1111/jev_localPython24GitHub stars
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.
Jev Practices & Patterns#community#jevbest#catalog-import
Community
rorshopping/jev-on-a-laptopPython
rorshopping/jev-on-a-laptopPython22GitHub stars
Offline evaluator that checks whether classifier confidences match reality and converts mistake costs into a deployable threshold, emitting an HTML report plus machine-readable threshold files.
Jev Practices & Patterns#community#jevbest#catalog-import
Community
rlaope/jevalPython
rlaope/jevalPython16GitHub stars
Vendor-neutral harness that tests whether Jev reranking improves a small RAG stack, using frozen hybrid BM25 plus BGE-M3 candidates and isolating retrieval, rerank, and generation stages.
Jev Practices & Patterns#community#jevbest#catalog-import
Community
erendikmenn/jev-rag-benchmarkPython
erendikmenn/jev-rag-benchmarkPython37GitHub stars
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.
Jev Practices & Patterns#community#jevbest#catalog-import
Community
sgoedecke/system-onePython
sgoedecke/system-onePython39GitHub stars
Open replica of the typed decision API on openly available base models, answering choice, score, and true-probability questions in one forward pass with no decoding.
Jev Practices & Patterns#community#jevbest#catalog-import
Community
mithalouni/system-one-openPython
mithalouni/system-one-openPythonFrom the Jev guides
Original write-ups on what this entry is for.
Most Jev accuracy problems are state problems. What belongs in a state, what is a pre-made judgment to leave out, and why naming the slice you ask about matters more than the wording.
Jev Practices & Patterns#state#design#primitives
Read article
A Noul returns one number and no separate confidence. How to read the value, pin its boundary with criteria, split compound questions, and set a per-action threshold that matches the cost of being wrong.
Jev Practices & Patterns#noul#primitives#confidence
Read article
A Score is a weighted position, not a ruler. Why the number can land between levels, why levels are judged separately, and how to combine several atomic scores in code instead of one crowded rubric.
Jev Practices & Patterns#score#primitives#scoring
Read article
From the Jev community
Posts from builders shipping with Jev right now.