Organization repository on GitHub
SearchJev: a fast and calibrated System-1 model for search agents (training code)
- Forks
- 1
- Last pushed
- Updated Oct 2026
Repo stats from the GitHub API, cached Oct 2026.
Jev directory
CommunityJev Repos & SDKs
SearchJev
Training code for SearchJev, a fast calibrated System-1 model for search agents: given a search state and a decision schema, it returns typed predictions over legal field values in one forward pass of a Qwen3.5 backbone.
- Stars
- 33
- Language
- Python
- License
- APACHE-2.0
- Reading time
- 2 min
- Published by
- Community
- Author
- EvoScientist
- Last verified
- Oct 2026
- Languages
- Python
- Capabilities
- Open source
Highlights
- Turns search decisions such as is this document relevant or is the evidence sufficient into schema-conditioned typed prediction.
- Returns a calibrated distribution over legal field values in a single forward pass, without autoregressive decoding.
- Everything is open: code, data (SearchDecision-Bench), model weights and complete training logs.
- Trains 0.8B and 4B paper models, and the same recipe also handles the 9B backbone: LoRA on Qwen3.5 with a first-token option-label readout.
- Loss combines soft-label cross-entropy with 0.5 x Brier; one temperature per question type is fitted on half the validation data.
Watch out
README reviewed 2026-10-17; the excerpt is truncated inside the repository-layout listing and the digest includes no install command or benchmark tables, so the paper's accuracy and calibration numbers could not be checked here.
Questions about this entry
Short answers drawn from the linked project or page.
A System-1 model for search agents, with its full training code in this repo. Given a search state and a decision schema, it returns a calibrated probability distribution over the legal values of every field in one forward pass, without autoregressive decoding.
The short structured choices a search agent makes: is this document relevant, is the evidence sufficient, where should the next query go, which link next. Schema-conditioned typed prediction replaces free-form generation for each of them.
The paper models SearchJev-0.8B and SearchJev-4B, and the README says the same recipe also trains the 9B backbone. Code, data, model weights and complete training logs are all open, with weights at huggingface.co/SearchJev.
LoRA on a Qwen3.5 backbone with a restricted first-token option-label readout, soft-label cross-entropy plus 0.5 times the Brier score, on a single GPU or DDP. start_training.py runs prepare, train, calibrate, held-out eval and summary in one go.
One temperature per question type is fitted on half of the validation data, and the other half is held out for evaluation. The training loss combines soft-label cross-entropy with half of the Brier score.
It follows the System One pattern rather than calling Jev: decisions are typed, options are explicit, and outputs are calibrated probabilities for the legal values. The README describes 'JEV files' converted from the SearchDecision-Bench for training.
EvoScientist owns the repo: 32 stars and 1 fork, Apache-2.0 licensed, written in Python, created 2026-10-03 and last pushed 2026-10-04. The paper PDF and training logs are linked from the README, but the digest does not include benchmark tables.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
X: LinkedIn bookmarks search
X: 1,066 reaction videos, searchable
X: Intent-based search in a Chrome extension
X: Margin, bookmark search
More Jev resources like this
A System One layout optimizer for decision-model workloads: it reorders rows and fields so a prefix-caching backend reuses input computation, then returns per-record decisions in original order. Its bundled backend serves AutoTrust JEV-9B through vLLM.
Jev Repos & SDKs#community#github#python
Community
0814wdwd/solo_jevPython
0814wdwd/solo_jevPythonA Bangla-focused adaptation of Laya, the non-autoregressive typed-decision model: targeted data preparation, fine-tuning and evaluation for classification, routing and scoring on Bangla, Banglish and code-switched text, with the checkpoint still to come.
Jev Repos & SDKs#community#github#python
Community
shamspias/laya-banglaPython
shamspias/laya-banglaPythonA TypeScript-first toolkit billed as the standard library for System One models: ready-made classify, extract, pickOne, rank and verify methods, a service-agent runtime that runs conversations turn by turn, and a one-message order parser. A Python port is in preview and the project is in alpha.
Jev Repos & SDKs#community#github#typescript
Community
Engid/question-kitTypeScript
Engid/question-kitTypeScriptA Python library shaped like the re module that finds, replaces, splits and extracts text by meaning: loose finders propose spans and Jev judges each one with a probability. Every match stays a verbatim slice of the input with its offsets, so it can be wrong but never invented.
Jev Repos & SDKs#community#github#python
Community
nothans/ex-regexPython
nothans/ex-regexPythonAn open-source, locally executed decision engine in the spirit of Jev (TypeSafe AI): state becomes features, and Noul, Choice and Score heads return typed, calibrated probabilities without generating text. It ships a cheap-to-strong cascade and a full train-calibrate-evaluate-deploy cycle.
Jev Repos & SDKs#community#github#python
Community
slavabars/decision-cascadePython
slavabars/decision-cascadePythonAn open-source, non-autoregressive System One decision engine inspired by TypeSafe AI: it evaluates state context against typed questions in a single forward pass, returning probability distributions and calibrated confidence without generating text.
Jev Repos & SDKs#community#github#python
Community
Atulmishra22/jev-litePython
Atulmishra22/jev-litePythonFrom the Jev guides
Original write-ups on what this entry is for.
From the Jev AI model community
Posts from builders shipping with Jev right now.