TiktokenSharp

Due to the lack of a C# version of cl100k_base encoding (gpt-3.5-turbo), I have implemented a basic solution with encoding and decoding methods based on the official Rust implementation.

Currently, cl100k_base p50k_base has been implemented. Other encodings will be added in future submissions. If you encounter any issues or have questions, please feel free to submit them on the lssues."

If you want to use the ChatGPT C# library that integrates this repository and implements context-based conversation, please visit ChatGPTSharp.

Getting Started

TiktokenSharp is available as NuGet package.

using TiktokenSharp;

//use model name
TikToken tikToken = TikToken.EncodingForModel("gpt-3.5-turbo");
var i = tikToken.Encode("hello world"); //[15339, 1917]
var d = tikToken.Decode(i); //hello world

//use encoding name
TikToken tikToken = TikToken.GetEncoding("cl100k_base");
var i = tikToken.Encode("hello world"); //[15339, 1917]
var d = tikToken.Decode(i); //hello world

When using a new encoder for the first time, the required tiktoken files for the encoder will be downloaded from the internet. This may take some time. Once the download is successful, subsequent uses will not require downloading again. You can set TikToken.PBEFileDirectory before using the encoder to modify the storage path of the downloaded tiktoken files, or you can pre-download the files to avoid network issues causing download failures.

Why are the tiktoken files not integrated into the package? On one hand, this would make the package size larger. On the other hand, I want to stay as consistent as possible with OpenAI's official Python code.

Below are the file download links: p50k_base.tiktoken cl100k_base.tiktoken

Efficiency Comparison

I noticed that some users would like to get a comparison of efficiency. Here, I use SharpToken as the basic comparison, with the encoder cl100k_base, on the .Net 6.0 in Debug mode.

TiktokenSharp Version: 1.0.5
SharpToken Version: 1.0.28

CPU

Code：

const string kLongText = "King Lear, one of Shakespeare's darkest and most savage plays, tells the story of the foolish and Job-like Lear, who divides his kingdom, as he does his affections, according to vanity and whim. Lear’s failure as a father engulfs himself and his world in turmoil and tragedy.";

static async Task SpeedTiktokenSharp()
{
    TikToken tikToken = TikToken.GetEncoding("cl100k_base");
    Stopwatch stopwatch = new Stopwatch();
    stopwatch.Start();

    for (int i = 0; i < 10000; i++) 
    {
        var encoded = tikToken.Encode(kLongText);
        var decoded = tikToken.Decode(encoded);
    }

    stopwatch.Stop();
    TimeSpan timespan = stopwatch.Elapsed;
    double milliseconds = timespan.TotalMilliseconds;
    Console.WriteLine($"SpeedTiktokenSharp = {milliseconds} ms");
}

static async Task SpeedSharpToken()
{
    var encoding = GptEncoding.GetEncoding("cl100k_base");

    Stopwatch stopwatch = new Stopwatch();
    stopwatch.Start();   

    for (int i = 0; i < 10000; i++) 
    {
        var encoded = encoding.Encode(kLongText);
        var decoded = encoding.Decode(encoded);
    }

    stopwatch.Stop();
    TimeSpan timespan = stopwatch.Elapsed;
    double milliseconds = timespan.TotalMilliseconds;
    Console.WriteLine($"SpeedSharpToken = {milliseconds} ms");

}

TiktokenSharp is approximately 57% faster than SharpToken.

SpeedTiktokenSharp = 570.1206 ms
SpeedSharpToken = 1312.8812 ms

Memory

Image：

TiktokenSharp has approximately 26% less memory usage than SharpToken.

Update

1.0.6 20230625

Replace WebClient with HttpClient, add async methods.

1.0.5 20230508

New support for .Net Standard 2.0 has been added, making TiktokenSharp usable in the .Net Framework.

1.0.4 20230424

Add method TikToken.GetEncoding(encodingName).

1.0.3 20230321

GetEncodingSetting now supports the model of gpt-4 and also allows for encoding names to be directly passed in.

1.0.2 20230317

add method TikToken.PBEFileDirectory to allow for custom storage directory of bpe files. the path needs to be set before TikToken.EncodingForModel().

1.0.1 20230313

p50k_base encoding algorithm that supports the text-davinci-003 model.

mHalo/TiktokenSharp