ndif-team

Neural network interpretability with nnsight and NDIF - tracing and intervening on model internals, debugging, remote execution, logit lens, activation patching, attribution patching, causal tracing, and steering

19 skills

Self-hosting NDIF, the server behind nnsight's remote=True - standing one up with docker run, compose, or the ndif CLI; deploying, pinning and sizing models; troubleshooting; and changing the server itself

4 skills

Architecture-agnostic interpretability with nnterp, a layer on nnsight that gives every transformer family the same module names and standard values - usage of the values, recipes that run unchanged across families, and extending it with a new family

3 skills