Jevlike is an open-source model that scores text options in one pass

Jevlike is a new open-source project on GitHub that copies the input and output format of TypeSafe's commercial Jev model. The repository went public on 16 September, and by the time Hacker News picked it up it had 305 stars and a score of 100 points on the front page.

TypeSafe's Jev, introduced last week, takes a piece of context and a list of N text options, and returns one probability for each option in a single forward pass, without writing an answer token by token. TypeSafe has not published how its model is trained. Jevlike, by Vinny LaRouge, is an independent starter with the same interface. It is MIT-licensed.

What it does

Each option becomes a query vector. That query attends over the context, producing one context vector per option, and a shared dot product turns each pair into one score. A softmax across the options gives the final probabilities. The default encoder learns byte embeddings from scratch. An optional path uses a frozen pretrained encoder from Hugging Face, with a small scorer head trained on top.

Training data is a JSONL file with context, options, and a zero-based label. The number of options can differ between rows, and rows must include every option the model will ever see at prediction time.

The numbers the author reports

The README states three results from local experiments that led to the starter, and is explicit that they are not this quickstart's output. On synthetic menus, the one-pass scorer reached about 98 percent accuracy. On target-disjoint Wikispeedia next-click data, a frozen Qwen2.5-0.5B encoder plus the scorer reached 26 percent, against about 8 percent for shuffled and random-encoder controls. A small from-scratch model on 40,000 clicks reached 29 percent. At eight options, one pass was about 100 times faster than a small local decoder forced to write 400 tokens.

The author states plainly that these numbers do not match commercial Jev, that TypeSafe's private training method has not been reproduced, and that the speed comparison used a small local decoder rather than a large commercial system. The Wikispeedia numbers assume target-disjoint splits, so a related record cannot leak into the test set.

The routing task, one pass per decision

The one-pass shape suits any routing task with a bounded list of choices: pick a menu item, a support queue, a category, a next click, or a controller button. A decoder can do these too, but it pays for each word it writes. Jevlike does one pass no matter how many options are on offer, and the byte-level default runs on CPU, MPS or CUDA.

The cost of that speed is that the model needs the full option list before prediction, and Jevlike's byte encoder is cheap but weak on language meaning. A frozen Hugging Face encoder helps at the cost of memory. For a real deployment the author recommends keeping related records inside the same split so near-duplicates cannot leak between training and evaluation, and running the shuffled-context control to confirm the model is doing more than guessing.

For teams already looking at Jev, this is the first public code that reproduces the interface, releases a working scorer, and states its evaluation method. It is a research starting point rather than a ready-to-use Jev copy, and it costs nothing.

Source

Jevlike on GitHub. Discussion on Hacker News. TypeSafe's Jev announcement sets the context for the interface Jevlike copies.