wtok 0.5

Count and truncate whisper text based on tokens

wtok: ttok for Whisper one sentence at a time

Count and truncate text based on tokens one sentence at a time

Background

Whisper models conditional distributions of a token given a sequence of past tokens

This tool can count tokens, using OpenAI's tiktoken library.

It can also truncate text to a specified number of tokens.

Installation

Install this tool using pip:

pip install wtok

Counting tokens

Provide text as arguments to this tool to count tokens:

wtok Hello world

2

You can also pipe text into the tool:

echo -n "Hello world" | wtok

2

Here the echo -n option prevents echo from adding a newline - without that you would get a token count of 3.

To pipe in text and then append extra tokens from arguments, use the -i - option:

echo -n "Hello world" | wtok more text -i -

6

Different models

By default, the tokenizer model for GPT-3.5 and GPT-4 is used.

To use the model for GPT-2 and GPT-3, add --model gpt2:

wtok boo Hello there this is -m gpt2

6

Compared to GPT-3.5:

wtok boo Hello there this is

5

Further model options are documented here.

Truncating text

Use the -t 10 or --truncate 10 option to truncate text to a specified number of tokens:

wtok This is too many tokens -t 3

This is too

Viewing tokens

The --encode option can be used to view the integer token IDs for the incoming text:

wtok Hello world --encode

9906 1917

The --decode method reverses this process:

wtok 9906 1917 --decode

Hello world

Add --tokens to either of these options to see a detailed breakdown of the tokens:

wtok Hello world --encode --tokens

[b'Hello', b' world']

Development

To contribute to this tool, first checkout the code. Then create a new virtual environment:

cd wtok
python -m venv venv
source venv/bin/activate

Now install for editing:

pip install -e .