Do I need a Jev account to use this at all?

Only for the teacher form. Passing --teacher typesafe-ai/jev sends your inputs to a hosted teacher and needs an AI gateway key in the environment; its responses cache to .teacher/ so the same command resumes. The other two forms train from answers you already have and make no hosted call.

Is this a hosted classifier service I have to call over a network?

No. It needs Node 22 or later and runs locally - Node inference needs no network access. The browser path copies runtime assets plus the trained model into your public directory and infers in a worker, so you serve it over HTTPS or localhost from your own app.

Can I treat the recommended probability as a hard threshold?

Treat it as a starting point only. The SDK returns every prediction and applying a cutoff is your application's choice, and thresholdRecommendation can come back null on small datasets - which is what happens with the bundled quick-start data. Calibrate in the runtime you will actually deploy to.

Why does my second training run fail immediately?

Output directories must not already exist, so a repeat run into models/custom aborts rather than overwriting. Delete or rename the previous directory first. The related trap on the deploy side is that the bundled encoder lives inside the output directory, so copy the whole directory when you move a model.

How do I install it and train the first model?

npm install jimothy on Node 22 or later, then npx jimothy train with a task definition, a JSONL or JSON dataset, and an --out directory. The default MiniLM backend downloads and bundles its pretrained encoder during training, so the first run reaches the network even though inference later does not.

Should I start with TF-IDF or with MiniLM?

They trade accuracy for speed. On BANKING77, TF-IDF trained in 37.6 s and predicted in 0.023 ms for 82.20% accuracy, while MiniLM took 7 min 26 s and 1.72 ms for 92.37%. Pass --backend tfidf for a fast baseline or --encoder to reuse MiniLM assets offline.

What does serving it in a browser actually require?

Run npx jimothy prepare-browser to write runtime assets into your public directory, copy the complete trained model under public/models, and serve over HTTPS or localhost. Inference runs in a worker: TF-IDF in JavaScript, MiniLM as q8 WASM by default, and FP16 WebGPU only with a matching calibrated export.

CommunityJev Demos & Experiments

Jimothy

A Node package that distils Jev-style typed answers into small local classifiers, trained either from saved labels or from a Jev teacher, then served in Node or a browser worker.

See highlights

Open github.com

Stars: 75
Language: TypeScript
License: MIT
Published by: Community
Author: AndrewPrifer
Last verified: Oct 2026

AndrewPrifer/jimothy

User repository on GitHub

View on GitHub

Forks: 3
Last pushed: Updated Sep 2026

Repo stats from the GitHub API, cached Oct 2026.Homepage

About this project

Train small, insanely fast local classifiers from Jev-compatible examples. Run locally in your browser or Node.js. Requires Node.js 22 or later. Install and train a support-ticket classifier using the included examples: This uses the default MiniLM backend, which downloads and bundles its pretrained encoder during training. Keep the entire output directory together when moving or deploying a model.

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

Tagscommunity jevbest

Works forJev for Classification Jev for Tool calling

Used inJev for Software & DevTools

Languages: TypeScript

Highlights

Three training forms: saved answers, split inputs and outputs joined by id, or a teacher model that labels your inputs.
Training tunes regularization, calibrates probabilities and recommends a cutoff; the SDK still returns every prediction.
TF-IDF trains about 12x faster and predicts about 75x faster; MiniLM scored 92.37% against 82.20% on BANKING77.
Browser inference runs in a worker: TF-IDF in JavaScript, MiniLM as q8 WASM or FP16 WebGPU.

Watch out

Accuracy figures are the project's own, measured on BANKING77 and a 300-sample email set on one machine, and the README states outright that results depend on the task, hardware, runtime and caches. MiniLM probabilities differ between runtimes, so a cutoff that works in Node can move in a browser; the README explicitly says to validate cutoffs in your deployment environment. Training on the default backend downloads and bundles a pretrained encoder, and an entire output directory has to stay together to move or deploy a model. Only the teacher form calls a hosted model. The captured README stops after the browser guide, so encoder preparation options and the full measurement methodology were not visible. 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-browser-skill

Standalone CLI skill that drives isolated Playwright Chromium with Jev choosing scoped in-page actions, defaulting to Vercel AI Gateway with TypeSafe as an option.