Here is a code release for the 2023 NeurIPS paper "How does GPT-2 compute greater-than?: Interpreting mathematical abilities in a pre-trained language model".

The code release is structured as follows:

• circuit_discovery.py reproduces the circuit discovery and semantics assignment process
• big_ds_experiments.py reproduces the experiments run on the larger, 10,000 element dataset
• neuron_investigations.py reproduces the neuron-level experiments
• sequence_generalization.py reproduces the generalization experiments

The aforementioned files will cache files (in paper-cache) that can be used to generate plots using these scripts:

• circuit_discovery_plotting.py
• pca_plots.py
• neuron_plots.py
• appendix_plots.py

In addition, we include three useful files in the cache folder (indices of the relevant logits, the nouns used in our template, and the order of MLP10 neurons, which otherwise takes a long time to compute). Finally, we include two utility files, utils.py and color_utils.py (for plotting).

Most of these experiments started as exploratory VSCode notebooks, but can be run just as easily as Python scripts, and will produce all necessary output. We also include a few smaller notebooks that correspond to discussions with reviewers, and didn't fit in neatly with the rest of our experiments:

• random_circuit_ablation.py: allows you to try ablating random circuits, as opposed to the one we found
• random_years.py: tests GPT-2's responses to random sequences of years from the same century
• topk_years.py: tests the degree to which GPT-2's top-k YY predictions are correct.

Unfortunately, using the rust-circuit library to work with gpt2-small is not easy. To run the code, follow these steps:

1. Compile rust-circuit, following the instructions there given; note that this requires clang and rust. The repo instructs you to install maturin; be sure to install 0.14.x (we used 0.14.7), as newer versions do not work.
2. Install this project's requirements via the provided requirements file pip install -r requirements.txt
3. Download the gpt2-small model files from this link. Extract them to a folder called ../rrfs/tensor_db.

Our paper is available on ArXiv and, hopefully sometime soon, on the NeurIPS website. You can cite it like so:

@inproceedings{
hanna2023how,
title={How does {GPT}-2 compute greater-than?: Interpreting mathematical abilities in a pre-trained language model},
author={Michael Hanna and Ollie Liu and Alexandre Variengien},
booktitle={Thirty-seventh Conference on Neural Information Processing Systems},
year={2023},
url={https://openreview.net/forum?id=p4PckNQR8k}
}