github.com via Hacker News

Jevlike Ships Open-Source Option Scorer with ~100x Speedup

TL;DR

  • Jevlike returns one probability per option in a single forward pass and reports being about 100x faster than a small decoder forced to write 400 tokens.
  • With a frozen Qwen2.5-0.5B encoder, the model reached 26% on Wikispeedia against roughly 8% for shuffled and random-encoder controls.
  • The README is explicit that this is a research starter, not a copy of TypeSafe's Jev, and equal quality with Jev has not been demonstrated.

Jevlike, a new open-source starter model on GitHub, takes a piece of text and a list of N options and returns one probability per option in a single forward pass. The README claims one pass was about "100 times faster than a small decoder forced to write 400 tokens." It is pitched as a research replica of the mechanism behind TypeSafe's commercial Jev, not the trained model itself.

On synthetic menus the authors report the model "reached about 98% accuracy on synthetic menus." On real Wikispeedia click data, a version wrapped around a frozen Qwen2.5-0.5B encoder "reached 26%, against about 8% for shuffled and random-encoder controls," while a small model trained from scratch "reached 29%" on 40,000 clicks. Each option becomes a query vector that attends over the context tokens, and a softmax over those attention scores yields the per-option probability.

The README is careful about what has and has not been shown: "This is a research starter, not a copy of Jev," the authors write, and note the numbers come from "local experiments, not this quickstart run." It lands amid a heavy run of open-source releases our tracker has logged over the last three months. Code is MIT-licensed; downloaded datasets and pretrained encoder weights keep their own terms.