JevMade Sign in
← Back to experiments

Benchmarks & research

jev-chat

Source screenshot of jev-chat
SOURCE SCREENSHOTFull screenshot ↗

What it does

jev-chat constructs chatbot replies by repeatedly turning Jev probabilities into hierarchical speculative decoding choices.

How you can use it

You can test this project with a developer using a TypeSafe API key. An API key is a secret software code that links your project to your service account. The app asks the decision model to choose pieces of text, like common phrases or individual words, and stitches those pieces together into a full reply.

First, write down sample questions and complete candidate answers for your project. This experiment is research into an unusual way to write text rather than a finished product. Generated sentences often get stuck repeating words or making math errors. Picking one full answer from your prepared list produces much more reliable results.