The Language Interpretability Tool (LIT) is a visual, interactive model-understanding tool for ML models, focusing on NLP use-cases. 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, attention, 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 installed via pip, or can be built from source. Building from source is necessary if you wish to update any of the front-end or core back-end code.
Download the repo and set up a Python environment:
git clone https://github.com/PAIR-code/lit.git ~/lit
# Set up Python environment
cd ~/lit
conda env create -f environment.yml
conda activate lit-nlp
conda install cudnn cupti # optional, for GPU support
conda install -c pytorch pytorch # optional, for PyTorch
# Build the frontend
pushd lit_nlp; yarn && yarn build; popdNote: if you see an error running yarn on Ubuntu/Debian, be sure you have the correct version installed.
pip install lit-nlpThe pip installation will install all necessary prerequisite packages for use of the core LIT package. It also installs the code to run our demo examples. It does not install the prerequisites for those demos, so you need to install those yourself if you wish to run the demos. See environment.yml for the list of all packages needed for running the demos.
Explore a collection of hosted demos on the LIT website demos page.
Colab notebooks showing the use of LIT inside of notebooks can be found at lit_nlp/examples/notebooks. A simple example can be viewed here.
To explore classification and regression models tasks from the popular GLUE benchmark:
python -m lit_nlp.examples.glue_demo --port=5432 --quickstartNavigate to http://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.
To explore predictions from a pretrained language model (BERT or GPT-2), run:
python -m lit_nlp.examples.lm_demo --models=bert-base-uncased \
--port=5432And navigate to http://localhost:5432 for the UI.
A simple colab demo can be found here. Just run all the cells to see LIT on an example classification model right in the notebook.
See docker.md for instructions on running LIT as a containerized web app. This is the approach we take for our website demos.
See lit_nlp/examples. Run similarly to the above:
python -m lit_nlp.examples.<example_name> --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
DatasetAPI - Write a model wrapper which follows the
ModelAPI - 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 against that branch. The main branch is for stable releases, and it is expected that the dev branch will always be ahead of main in terms of commits.
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'd love your suggestions and feedback - drop us a line in the issues.