lit

Why we included this project

LIT (the Learning Interpretability Tool) is a visual, browser-based way to interrogate a machine learning model instead of trusting it on faith. You run it as a local server or inside a Jupyter or Colab notebook, load a text, image, or tabular model, and then ask pointed questions: where it makes mistakes, why it chose a particular prediction, and whether the answer shifts when you edit the input. For explanations it offers salience maps on individual predictions, slicing and binning for aggregate behavior, counterfactual generation, and a side-by-side mode that compares two models on the same examples. It works with TensorFlow and PyTorch without touching your training code, and its plug-in design extends to new model types, so it slots into an existing ML workflow more easily than most interpretability research tools. That makes it practical for debugging a model before release and for getting a concrete feel for what the model has actually learned.