[go: up one dir, main page]

Skip to content

sctg-development/sentencepiece-js

Repository files navigation

Javascript wrapper for the sentencepiece library

Build React App Publish to npmjs registry

Browser Demo

You can see Sentencepiece-js in action for counting and displaying tokens using the Meta Llama 3.1 tokenizer model on GitHub Pages: https://sctg-development.github.io/sentencepiece-js/. All computations are performed in your browser, and no data is sent to the server. To display the tokens, click on the tokens link.

This simple React app is located in the tokenCount directory of this repository. It is built with React 18, Vite, and the Fluent UI v9 framework.

Build

Sentencepiece is compiled to webassembly using emscripten.

To rebuild this project

npm install

git clone --recurse-submodules  https://github.com/sctg-development/sentencepiece-js.git

npm run build

Use

To use this tool in nodejs, you can use the following code:

const { SentencePieceProcessor, cleanText } = require("../dist");
const ROOT = require('app-root-path')

async function main() {

    let text = "I am still waiting on my card?"
    let cleaned = cleanText(text)

    let spp = new SentencePieceProcessor()
    await spp.load(`${ROOT}/test/30k-clean.model`)
    let ids = spp.encodeIds(cleaned)
    console.log(ids)
    let str = spp.decodeIds(ids) // list ids->number
    console.log(str)

    let pieces = spp.encodePieces(cleaned) // list tokens->string
    console.log(pieces)
}
main()

In the browser, you can use the following code (see the tokenCount directory for a full example):

import { SentencePieceProcessor, cleanText, llama_3_1_tokeniser_b64 } from "@sctg/sentencepiece-js";

// built in models: llama_3_1_tokeniser_b64, clean_30k_b64, smart_b64
async function main() {

    let text = "I am still waiting on my card?"
    let cleaned = cleanText(text)

    let spp = new SentencePieceProcessor()
    await spp.loadFromB64StringModel(llama_3_1_tokeniser_b64);
    let ids = spp.encodeIds(cleaned)
    console.log(ids)
    let str = spp.decodeIds(ids) // list ids->number
    console.log(str)

    let pieces = spp.encodePieces(cleaned) // list tokens->string
    console.log(pieces)
}
main()

See https://github.com/sctg-development/ai-outlook/blob/HEAD/src/aipane/aipane.ts#L12-L34 for an example of how to use this in a react app.
Look also at webpack.config.js for the configuration of the webpack bundler.

  • devilyouwei updated this repo to make this module support the js require keyword and added the using example.
  • 2023-1-10, devilyouwei added encodePieces.
  • original author: https://github.com/JanKaul/sentencepiece