Skip to content

Token level visualization tools for large language models

Notifications You must be signed in to change notification settings

FateScript/token_visualizer

Repository files navigation

Introduction

token visualizer is a token-level visualization tool to visualize LLM.

Quick start

Installation

Install from source

Run the following command to install the package.

git clone git@github.com:FateScript/token_visualizer.git
cd token_visualizer
pip install -v -e .  # or  python3 setup.py develop

Check installation

If you could see the version of token_visualizer by running

python3 -c "import token_visualizer; print(token_visualizer.__file__)"

Visualization demo

Inference

Start demo

Run the following command to start inference visualizer.

python3 visualizer.py

The command will start a OpenAIProxy model, to use it without exception, user should fill in the value of BASE_URL and OPENAI_KEY.

token_visualizer also support OpenAIModel and HuggingFace TransformerModel in models.py, feel free to modify the code.

Demo gif

After inputing your prompt, you will see the large language model's answer and the answer's visualization result.

The redder the color of the token, the lower the corresponding probability. The greener the color of the token, the higher the corresponding probability.

Perplexity

Start demo

Run the following command to start perplexity visualizer, then click the ppl tab.

python3 visualizer.py

Demo gif

After inputing your text, you will see the perplexity and visualization result of the text.

Tokenizer

Start demo

Run the following command to start interactive tokenizer encoding web demo.

python3 visual_tokenizer.py

Demo gif

User could select tokenizer to interacte with and text to encode. For speical string

TODO

  • Support ppl visualization.
  • Select transformers/openai/TGI with cli.
  • Support OpenAI tokenizer visualization.
  • Support TGI inference visualization.
  • Support multi-turn chat visualization.
  • Support dark mode.

Related projects/websites

Acknowledgement

About

Token level visualization tools for large language models

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published