This repository includes the accompanying code for the paper "Transformer Feed-Forward Layers Build Predictions by Promoting Concepts in the Vocabulary Space ". Mor Geva*, Avi Caciularu*, Kevin Ro Wang, and Yoav Goldberg.

In this repository, we provide the main scripts and notebooks used to reproduce our experiments.

The repository contains:

• Code for generating the dataframe, based on the WikiText corpus, which served us during the experiments (in data_generation dir).
• Code for generating the values' projections (in values_projections dir).
• Code for generating the values' clusters (in clustering dir).
• Code for reproducing the toxic language suppression results (in toxic_language_suppression dir).
• Code for reproducing the early exit results (in early_exit dir).

For additional data or code please contact Avi, Mor or Kevin.

If you find our work useful, please cite the paper as:

@article{geva2022transformer,
      title={Transformer Feed-Forward Layers Build Predictions by Promoting Concepts in the Vocabulary Space},
      author={Geva, Mor and Caciularu, Avi and Wang, Kevin Ro and Goldberg, Yoav},
      journal={arXiv preprint arXiv:2203.14680},
      year={2022},
}