Create a language model with Next-Token Prediction for high-quality predictions in text, phrases, or even pixels as an alternative to decoder-only models like GPT or Mistral. Simple usage involves installing npm i next-token-prediction and setting up a training directory in your project. After training, the model can predict words, phrases, and completion suggestions. For advanced users, training data can be provided or generated from text files. The goal is to provide a top-quality token prediction library for auto-completion, spell checking, and search/lookup, aiming to simplify methodologies and create a free/open LLM in JavaScript. View examples and inspiration from 3Blue1Brown’s YouTube video.
https://github.com/bennyschmidt/next-token-prediction