Unofficial bindings / ports in other languages

[hauntsaninja]

The following projects are not maintained by OpenAI. I cannot vouch that any of them are correct or safe to use. Use at your own risk.

Note that if a tokeniser fails to exactly match tiktoken's behaviour, you may get worse results when sampling from models, with no warning.

Javascript

• https://github.com/dqbd/tiktoken
• https://github.com/ceifa/tiktoken-node
• https://github.com/niieani/gpt-tokenizer
• The gpt-3-encoder package will work for most GPT-3 models. However, it will often appear to work for Codex or GPT-3.5 while actually being out of distribution, and will not at all work for GPT-4 or embeddings models.

Rust

• https://github.com/zurawiki/tiktoken-rs

Java

• https://github.com/eisber/tiktoken
• https://github.com/knuddelsgmbh/jtokkit

Ruby

• https://github.com/volition-co/tiktoken

C#

• https://github.com/dmitry-brazhenko/SharpToken

Go

• https://github.com/tiktoken-go/tokenizer
• https://github.com/pkoukk/tiktoken-go

PHP

• https://github.com/danny50610/bpe-tokeniser

Kotlin

• https://github.com/aallam/ktoken

Thanks to everyone for building useful things!

I'm happy to link to other projects in this comment.

[bluescreen10]

👋,

I built a port for go that you can find in the link below

https://github.com/tiktoken-go/tokenizer

[fang2hou]

I am currently using a another port in Go.
https://github.com/pkoukk/tiktoken-go

[rex-remind101]

Hello @hauntsaninja , I was looking at https://github.com/openai/tiktoken/blob/main/src/lib.rs and it appears to be written in Rust. Could this be open sourced into a crate of its own?

[hauntsaninja]

See the FAQ #98

[danielcompton]

@hauntsaninja would it be possible to publish the full test suite publicly? That would make it easier to tell whether a given implementation matches (or is close to) the official implementation.

[niieani]

Here's a pure JavaScript / TypeScript port of tiktoken: https://github.com/niieani/gpt-tokenizer
Playground online: https://gpt-tokenizer.dev

[niieani]

@shylockWu they're not incorrect. You've set gpt-tokenizer to tokenize using GPT-3.5/GPT-4 encoding, whereas the official openAI token calculator uses the older GPT-3. If you switch the playground to use the older model, you'll get the same result.

[danny50610]

👋

I ported a version of PHP, link here

https://github.com/danny50610/bpe-tokeniser

[aallam]

I have built and published a port for Kotlin: https://github.com/aallam/ktoken :)

[Gabriella439]

Pure Haskell implementation of tiktoken: https://hackage.haskell.org/package/tiktoken