eap-ig-faithfulness

This repo contains the code for the paper "Automatic Circuit Finding and Faithfulness" (COLM 2024; Arxiv link). The code for manipulating model graphs and performing EAP/-IG is contained in a submodule, EAP-IG; make sure to pull the commit / submodule here, as I've updated the EAP-IG module since releasing this experimental code. After pulling that submodule, you can install it by calling pip install -e . in its directory. A Conda environment that can be used to run these experiments is contained in environment.yml.

You can replicate the experiments in the paper as follows:

• For each task in ['ioi', 'greater-than', 'sva', 'gender-bias', 'fact-retrieval-comma', 'hypernymy-comma']:
  • first, run pareto.py to collect results for EAP/EAP-IG. You must specify the model name (--model), task (--task), task metric (either logit_diff or prob_diff, --metric), and --batch_size. If the model is large, you might instead want to run pareto_big.py, which provides options to separately set the --eval_batch_size.
  • then, run get_real_values.py, with similar options. This can be slow, as this script performs the actual activation patching metric change values with which you will compare the EAP/-IG estimates / circuits.
• Then, you will be ready to recreate most of the figures in the paper. Once you do so, the figures will be generated as a .png or .pdf file in the corresponding subfolder of the relevant directory in results.
  • To recreate Figure 1, you can run results/first_figure/first_figure.py.
  • To recreate Figure 3, you can run results/real_pareto_combined/plot_real_pareto_normalized.py.
  • To recreate Figure 4, you can run results/real_rank/compare_real.py and then results/real_rank/plot_overlap.py.
• To replicate the overlap experiments (Figure 5), you can run overlap_heatmaps.py as well as all-cross-task-faithfulness.py. Then, run results/cross-task/make_all_heatmaps.py; if you want recall heatmaps, as in Appendix I, run results/cross-task/recall_heatmaps.py
• The figures in Appendix can be replicated as follows.
  • For Figure 6, run results/manual_overlap/ioi_overlap.py and results/manual_overlap/greater_overlap.py
  • For Figure 7, run test_ig_iterations.py, taking care to specify the --model and --task of interest, as well as the --attribution_metric and the --eval_metric. You can generate the figure using results/ig_test/plot_ig_test.py.
  • For Figure 8, run results/real_rank/plot_kendall.py.
  • For Figure 9, run results/real_rank/node_edge_overlap_plot.py.
  • For Figure 10, run results/pareto/plot_pareto_normalized_single.py.
  • For Figure 11, run the larger models as described in the first step, and then again run results/real_pareto_combined/plot_real_pareto_normalized.py.

The data used in this paper is available in the data folder.