A research computing project enabling transparent and reproducible experiments on large-scale AI systems.
A Python library for interpreting and manipulating the internals of deep learning models. Access activations, modify them to study causal effects, compute gradients, and batch interventions efficiently.
Documentation | GitHub | Paper
pip install nnsight
The National Deep Inference Fabric — a nationwide research computing infrastructure enabling scientists and students to perform transparent experiments on running AI models without local GPU resources.
pip install ndif
A unified interface for mechanistic interpretability of transformers. Built on NNsight, it provides standardized naming conventions across all transformer architectures with built-in interventions like logit lens, patchscope, and activation steering.
Documentation | GitHub | Paper
pip install nnterp
A UI for doing exploratory analysis on open source AI models by applying interpretability techniques. Leverages both NNsight and NDIF to provide an interactive environment for exploring LLM internals.
