About nnterp¶
nnterp gives every transformer family the same module names and the same standard values, on top of nnsight. It is developed by the NDIF team and maintained at github.com/ndif-team/nnterp. The encyclopedia is one page per supported family, built from the family modules themselves. The skills that teach an agent nnterp are in ndif-team/skills.
nnterp 1.x was written by Clément Dumas and described in nnterp: A Standardized Interface for Mechanistic Interpretability of Transformers. nnterp 2 is a rebuild on nnsight 0.8: the names are nnsight aliases, the values are properties on the envoys, and each family is one module that says how its own names map onto the standard ones.
Citation¶
@misc{dumas2025nnterp,
title={nnterp: A Standardized Interface for Mechanistic Interpretability of Transformers},
author={Cl{\'e}ment Dumas},
year={2025},
eprint={2511.14465},
archivePrefix={arXiv},
primaryClass={cs.LG},
url={https://arxiv.org/abs/2511.14465},
}
Found a bug?¶
Reduce it to a minimal example first. If the same code written with nnsight's TransformersModel
on the native module names also fails, the bug is in nnsight: open it on the
nnsight tracker. Otherwise open an issue on the
nnterp tracker.