The Learning Interpretability Tool (🔥LIT, formerly known as the Language Interpretability Tool) is a visual, interactive ML model-understanding tool that supports text, image, and tabular data. It can be run as a standalone server, or inside of notebook environments such as Colab, Jupyter, and Google Cloud Vertex AI notebooks.
LIT is built to answer questions such as:
- What kind of examples does my model perform poorly on?
- Why did my model make this prediction? Can this prediction be attributed to adversarial behavior, or to undesirable priors in the training set?
- Does my model behave consistently if I change things like textual style, verb tense, or pronoun gender?
LIT supports a variety of debugging workflows through a browser-based UI. Features include:
- Local explanations via salience maps and rich visualization of model predictions.
- Aggregate analysis including custom metrics, slicing and binning, and visualization of embedding spaces.
- Counterfactual generation via manual edits or generator plug-ins to dynamically create and evaluate new examples.
- Side-by-side mode to compare two or more models, or one model on a pair of examples.
- Highly extensible to new model types, including classification, regression, span labeling, seq2seq, and language modeling. Supports multi-head models and multiple input features out of the box.
- Framework-agnostic and compatible with TensorFlow, PyTorch, and more.
LIT has a website with live demos, tutorials, a setup guide and more.
Stay up to date on LIT by joining the lit-announcements mailing list.
For a broader overview, check out our paper and the user guide.
LIT can be run via container image, installed via pip
or built from source.
Building from source is necessary if you update any of the front-end or core
back-end code.
Build the image using docker
or podman
:
git clone https://github.com/PAIR-code/lit.git && cd lit
docker build --file Dockerfile --tag lit-nlp .
See the advanced guide for detailed instructions on using the default LIT Docker image, running LIT as a containerized web app in different scenarios, and how to creating your own LIT images.
pip install lit-nlp
The pip
installation will install all necessary prerequisite packages for use
of the core LIT package.
It does not install the prerequisites for the provided demos, so you need to install those yourself. See requirements_examples.txt for the list of packages required to run the demos.
Clone the repo:
git clone https://github.com/PAIR-code/lit.git && cd lit
Note: be sure you are running Python 3.9+. If you have a different version on
your system, use the conda
instructions below to set up a Python 3.9
environment.
Set up a Python environment with venv
:
python -m venv .venv
source .venv/bin/activate
Or set up a Python environment using conda
:
conda create --name lit-nlp
conda activate lit-nlp
conda install python=3.9
conda install pip
Once you have the environment, install LIT's dependencies:
python -m pip install -r requirements.txt
python -m pip install cudnn cupti # optional, for GPU support
python -m pip install torch # optional, for PyTorch
# Build the frontend
(cd lit_nlp; yarn && yarn build)
Note: Use the -r requirements.txt
option to install every dependency required
for the LIT library, its test suite, and the built-in examples. You can also
install subsets of these using the -r requirements_core.txt
(core library),
-r requirements_test.txt
(test suite), -r requirements_examples.txt
(examples), and/or any combination thereof.
Note: if you see an error
running yarn
on Ubuntu/Debian, be sure you have the
correct version installed.
Explore a collection of hosted demos on the demos page.
To explore classification and regression models tasks from the popular GLUE benchmark:
python -m lit_nlp.examples.glue.demo --port=5432 --quickstart
Navigate to https://localhost:5432 to access the LIT UI.
Your default view will be a small BERT-based model fine-tuned on the Stanford Sentiment Treebank, but you can switch to STS-B or MultiNLI using the toolbar or the gear icon in the upper right.
And navigate to https://localhost:5432 for the UI.
### Notebook usage
Colab notebooks showing the use of LIT inside of notebooks can be found at
[lit_nlp/examples/notebooks](./lit_nlp/examples/notebooks).
We provide a simple
[Colab demo](https://colab.research.google.com/github/PAIR-code/lit/blob/main/lit_nlp/examples/notebooks/LIT_sentiment_classifier.ipynb).
Run all the cells to see LIT on an example classification model in the notebook.
### More Examples
See [lit_nlp/examples](./lit_nlp/examples). Most are run similarly to the
quickstart example above:
```sh
python -m lit_nlp.examples.<example_name>.demo --port=5432 [optional --args]
To learn about LIT's features, check out the user guide, or watch this video.
You can easily run LIT with your own model by creating a custom demo.py
launcher, similar to those in lit_nlp/examples. The
basic steps are:
- Write a data loader which follows the
Dataset
API - Write a model wrapper which follows the
Model
API - Pass models, datasets, and any additional components to the LIT server class
For a full walkthrough, see adding models and data.
LIT is easy to extend with new interpretability components, generators, and more, both on the frontend or the backend. See our documentation to get started.
To make code changes to LIT, please work off of the dev
branch and
create pull requests
(PRs) against that branch. The main
branch is for stable releases, and it is
expected that the dev
branch will always be ahead of main
.
Draft PRs are encouraged, especially for first-time contributors or contributors working on complex tasks (e.g., Google Summer of Code contributors). Please use these to communicate ideas and implementations with the LIT team, in addition to issues.
Prior to sending your PR or marking a Draft PR as "Ready for Review", please run the Python and TypeScript linters on your code to ensure compliance with Google's Python and TypeScript Style Guides.
# Run Pylint on your code using the following command from the root of this repo
pushd lit_nlp & pylint & popd
# Run ESLint on your code using the following command from the root of this repo
pushd lit_nlp & yarn lint & popd
If you use LIT as part of your work, please cite our EMNLP paper:
@misc{tenney2020language,
title={The Language Interpretability Tool: Extensible, Interactive Visualizations and Analysis for {NLP} Models},
author={Ian Tenney and James Wexler and Jasmijn Bastings and Tolga Bolukbasi and Andy Coenen and Sebastian Gehrmann and Ellen Jiang and Mahima Pushkarna and Carey Radebaugh and Emily Reif and Ann Yuan},
booktitle = "Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing: System Demonstrations",
year = "2020",
publisher = "Association for Computational Linguistics",
pages = "107--118",
url = "https://www.aclweb.org/anthology/2020.emnlp-demos.15",
}
This is not an official Google product.
LIT is a research project and under active development by a small team. There will be some bugs and rough edges, but we're releasing at an early stage because we think it's pretty useful already. We want LIT to be an open platform, not a walled garden, and we would love your suggestions and feedback - drop us a line in the issues.