Byte Tokenizer turns text into token IDs and turns those IDs back into text. Ordinary characters become raw UTF-8 bytes from 0 to 255, while optional special tokens start at 256 and match longest-first, so <|im_start|> is not split into a shorter <|im|> prefix. There is no vocabulary file, lowercase folding, or Unicode normalization, so case, combining marks, null bytes, and a leading BOM stay as encoded. Duplicate special tokens are ignored, and invalid UTF-8 bytes decode as U+FFFD.

Deno:

deno add npm:@neabyte/byte-tokenizer

npm:

npm install @neabyte/byte-tokenizer

CDN (jsDelivr/esm.sh):

<script type="module">
  import Tokenizer from 'https://cdn.jsdelivr.net/npm/@neabyte/byte-tokenizer/dist/index.mjs'
</script>

Or via esm.sh:

<script type="module">
  import Tokenizer from 'https://esm.sh/@neabyte/byte-tokenizer'
</script>

import Tokenizer from '@neabyte/byte-tokenizer'

// Raw bytes from 0 to 255
const tokenizer = new Tokenizer()

const rawIds = tokenizer.encode('Hello!')
// [72, 101, 108, 108, 111, 33]

const rawText = tokenizer.decode(rawIds)
// 'Hello!'

// Special tokens start at 256
const customTokenizer = new Tokenizer(['<|bos|>', '<|eos|>'])

const customIds = customTokenizer.encode('<|bos|>Hello!<|eos|>')
// [256, 72, 101, 108, 108, 111, 33, 257]

const customText = customTokenizer.decode(customIds)
// '<|bos|>Hello!<|eos|>'

new Tokenizer(specialToken?)

• specialToken <string[]>: (Optional) Special token strings. Defaults to []. Empty strings are rejected. Duplicate strings keep the first ID. IDs start at 256 and follow first-seen order.
• Throws: TypeError when the argument is not an array, or a token is empty or not a string.

tokenizer.encode(text)

• text <string>: Text to encode. Plain spans become UTF-8 byte IDs. A matching special token becomes one ID, and the longest match wins.
• Returns: <number[]>
• Throws: TypeError when text is not a string.

tokenizer.decode(identifier)

• identifier <ArrayLike<number>>: Token IDs, including typed arrays. IDs 0 to 255 become bytes. IDs 256 and above expand to their special-token text. A leading BOM is kept.
• Returns: <string>
• Throws: TypeError when the value is not array-like, or an ID is not an integer.
• Throws: RangeError when an ID is negative, -0, or outside the vocabulary.

npm run build

deno task check

deno task test

This project is licensed under the MIT License.