The fastest tokenizer for GPT-3.5 and GPT-4 inspired by Tiktoken.
$ dotnet add package TiktokenDue 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."
using Tiktoken;
var encoding = Encoding.ForModel("gpt-3.5-turbo");
var tokens = encoding.Encode("hello world"); // [15339, 1917]
var text = encoding.Decode(tokens); // hello world
var encoding = Encoding.Get("cl100k_base");
var tokens = encoding.Encode("hello world"); // [15339, 1917]
var text = encoding.Decode(tokens); // hello world
Once the download is successful, subsequent uses will not require downloading again. You can set 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.
<!--BENCHMARKS_START--> <!--BENCHMARKS_END-->TikToken.PBEFileDirectoryWhy 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
You can view the reports for each version here
BenchmarkDotNet=v0.13.5, OS=macOS Ventura 13.3.1 (a) (22E772610a) [Darwin 22.4.0]
Apple M1 Pro, 1 CPU, 10 logical and 10 physical cores
.NET SDK=7.0.203
[Host] : .NET 7.0.5 (7.0.523.17405), Arm64 RyuJIT AdvSIMD
Job-IGIIQY : .NET 7.0.5 (7.0.523.17405), Arm64 RyuJIT AdvSIMD DEBUG
BuildConfiguration=Debug
| Method | Data | Mean | Error | StdDev | Ratio | RatioSD | Gen0 | Gen1 | Allocated | Alloc Ratio |
|---|---|---|---|---|---|---|---|---|---|---|
| SharpTokenV1_0_28 | Hello, World! | 3.334 μs | 0.0497 μs | 0.0465 μs | 1.00 | 0.00 | 0.6752 | 0.0038 | 4.14 KB | 1.00 |
| TiktokenSharpV1_0_5 | Hello, World! | 6.640 μs | 0.0173 μs | 0.0162 μs | 1.99 | 0.03 | 2.1820 | 0.0458 | 13.41 KB | 3.24 |
| Tiktoken | Hello, World! | 8.957 μs | 0.0392 μs | 0.0327 μs | 2.68 | 0.04 | 2.1820 | 0.0458 | 13.41 KB | 3.24 |
| SharpTokenV1_0_28 | King(...)edy. [275] | 62.740 μs | 0.2538 μs | 0.2119 μs | 1.00 | 0.00 | 8.5449 | 0.4883 | 52.89 KB | 1.00 |
| TiktokenSharpV1_0_5 | King(...)edy. [275] | 21.973 μs | 0.1803 μs | 0.1599 μs | 0.35 | 0.00 | 5.0964 | 0.3052 | 31.34 KB | 0.59 |
| Tiktoken | King(...)edy. [275] | 23.805 μs | 0.0975 μs | 0.0814 μs | 0.38 | 0.00 | 5.7678 | 0.3357 | 35.38 KB | 0.67 |