ecco

Why we included this project

Ecco is for people who work with transformer language models and want to see what is happening inside them, not just call them. It runs inside Jupyter notebooks and turns model internals into interactive visualizations: which input tokens drove a prediction, how output token probabilities shift layer by layer, and which neurons respond to particular text patterns. It works with GPT2, BERT, RoBERTa, T5, T0, and other Hugging Face PyTorch models, and it can load your own local checkpoints too. Integrated Gradients and Saliency are built in, so you can trace a prediction back to the words that actually mattered. The library deliberately skips training and fine-tuning, which keeps it focused and easy to pick up. That focus is what makes it useful whether you are debugging a model's behavior, writing about interpretability, or showing students how attention and layers work.