Due to the original repository being inactive for an extended period, we have created a fork here. This version includes updates tailored for the latest versions of gpt-o4
and o200k
encoder.
OpenAI's tiktoken in Go.
Tiktoken is a fast BPE tokeniser for use with OpenAI's models.
This is a port of the original tiktoken.
go get github.com/pkoukk/tiktoken-go
Tiktoken-go has the same cache mechanism as the original Tiktoken library.
You can set the cache directory by using the environment variable TIKTOKEN_CACHE_DIR.
Once this variable is set, tiktoken-go will use this directory to cache the token dictionary.
If you don't set this environment variable, tiktoken-go will download the dictionary each time you initialize an encoding for the first time.
If you don't want to use cache or download the dictionary each time, you can use alternative BPE loader.
Just call tiktoken.SetBpeLoader
before calling tiktoken.GetEncoding
or tiktoken.EncodingForModel
.
BpeLoader
is an interface, you can implement your own BPE loader by implementing this interface.
The offline BPE loader loads the BPE dictionary from embed files, it helps if you don't want to download the dictionary at runtime.
Due to the size of the BPE dictionary, this loader is in other project.
Include if you require this loader: tiktoken_loader
package main
import (
"fmt"
"github.com/pkoukk/tiktoken-go"
)
func main() {
text := "Hello, world!"
encoding := "cl100k_base"
// if you don't want download dictionary at runtime, you can use offline loader
// tiktoken.SetBpeLoader(tiktoken_loader.NewOfflineLoader())
tke, err := tiktoken.GetEncoding(encoding)
if err != nil {
err = fmt.Errorf("getEncoding: %v", err)
return
}
// encode
token := tke.Encode(text, nil, nil)
//tokens
fmt.Println((token))
// num_tokens
fmt.Println(len(token))
}
package main
import (
"fmt"
"github.com/pkoukk/tiktoken-go"
)
func main() {
text := "Hello, world!"
encoding := "gpt-3.5-turbo"
tkm, err := tiktoken.EncodingForModel(encoding)
if err != nil {
err = fmt.Errorf("getEncoding: %v", err)
return
}
// encode
token := tkm.Encode(text, nil, nil)
// tokens
fmt.Println(token)
// num_tokens
fmt.Println(len(token))
}
Below is an example function for counting tokens for messages passed to gpt-3.5-turbo or gpt-4.
The following code was written based on openai-cookbook examples at Wednesday, 28 June 2023
.
Please note that the token calculation method for the message may change at any time, so this code may not necessarily be applicable in the future.
If you need accurate calculation, please refer to the official documentation.
If you find that this code is no longer applicable, please feel free to submit a PR or Issue.
package main
import (
"fmt"
"github.com/pkoukk/tiktoken-go"
"github.com/sashabaranov/go-openai"
)
// OpenAI Cookbook: https://github.com/openai/openai-cookbook/blob/main/examples/How_to_count_tokens_with_tiktoken.ipynb
func NumTokensFromMessages(messages []openai.ChatCompletionMessage, model string) (numTokens int) {
tkm, err := tiktoken.EncodingForModel(model)
if err != nil {
err = fmt.Errorf("encoding for model: %v", err)
log.Println(err)
return
}
var tokensPerMessage, tokensPerName int
switch model {
case "gpt-3.5-turbo-0613",
"gpt-3.5-turbo-16k-0613",
"gpt-4-0314",
"gpt-4-32k-0314",
"gpt-4-0613",
"gpt-4-32k-0613":
tokensPerMessage = 3
tokensPerName = 1
case "gpt-3.5-turbo-0301":
tokensPerMessage = 4 // every message follows <|start|>{role/name}\n{content}<|end|>\n
tokensPerName = -1 // if there's a name, the role is omitted
default:
if strings.Contains(model, "gpt-3.5-turbo") {
log.Println("warning: gpt-3.5-turbo may update over time. Returning num tokens assuming gpt-3.5-turbo-0613.")
return NumTokensFromMessages(messages, "gpt-3.5-turbo-0613")
} else if strings.Contains(model, "gpt-4") {
log.Println("warning: gpt-4 may update over time. Returning num tokens assuming gpt-4-0613.")
return NumTokensFromMessages(messages, "gpt-4-0613")
} else {
err = fmt.Errorf("num_tokens_from_messages() is not implemented for model %s. See https://github.com/openai/openai-python/blob/main/chatml.md for information on how messages are converted to tokens.", model)
log.Println(err)
return
}
}
for _, message := range messages {
numTokens += tokensPerMessage
numTokens += len(tkm.Encode(message.Content, nil, nil))
numTokens += len(tkm.Encode(message.Role, nil, nil))
numTokens += len(tkm.Encode(message.Name, nil, nil))
if message.Name != "" {
numTokens += tokensPerName
}
}
numTokens += 3 // every reply is primed with <|start|>assistant<|message|>
return numTokens
}
Encoding name | OpenAI models |
---|---|
cl100k_base |
gpt-4 , gpt-3.5-turbo , text-embedding-ada-002 |
p50k_base |
Codex models, text-davinci-002 , text-davinci-003 |
r50k_base (or gpt2 ) |
GPT-3 models like davinci |
Model name | OpenAI models |
---|---|
gpt-4-* | cl100k_base |
gpt-3.5-turbo-* | cl100k_base |
gpt-4 | cl100k_base |
gpt-3.5-turbo | cl100k_base |
text-davinci-003 | p50k_base |
text-davinci-002 | p50k_base |
text-davinci-001 | r50k_base |
text-curie-001 | r50k_base |
text-babbage-001 | r50k_base |
text-ada-001 | r50k_base |
davinci | r50k_base |
curie | r50k_base |
babbage | r50k_base |
ada | r50k_base |
code-davinci-002 | p50k_base |
code-davinci-001 | p50k_base |
code-cushman-002 | p50k_base |
code-cushman-001 | p50k_base |
davinci-codex | p50k_base |
cushman-codex | p50k_base |
text-davinci-edit-001 | p50k_edit |
code-davinci-edit-001 | p50k_edit |
text-embedding-ada-002 | cl100k_base |
text-similarity-davinci-001 | r50k_base |
text-similarity-curie-001 | r50k_base |
text-similarity-babbage-001 | r50k_base |
text-similarity-ada-001 | r50k_base |
text-search-davinci-doc-001 | r50k_base |
text-search-curie-doc-001 | r50k_base |
text-search-babbage-doc-001 | r50k_base |
text-search-ada-doc-001 | r50k_base |
code-search-babbage-code-001 | r50k_base |
code-search-ada-code-001 | r50k_base |
gpt2 | gpt2 |
you can run test in test folder
compare with original tiktoken
you can run benchmark in test folder
name | time/op | os | cpu | text | times |
---|---|---|---|---|---|
tiktoken-go | 8795ns | macOS 13.2 | Apple M1 | UDHR | 100000 |
tiktoken | 8838ns | macOS 13.2 | Apple M1 | UDHR | 100000 |
It looks like the performance is almost the same.
Maybe the difference is due to the difference in the performance of the machine.
Or maybe my benchmark method is not appropriate.
If you have better benchmark method or if you want add your benchmark result, please feel free to submit a PR.