README.md

Memorisation localisation for natural language classification tasks

0. Prepare by installing the conda environment using conda env create --file=env.yaml, which will install the environment memloc in python 3.9.
1. Add folders logs, checkpoints and subfolders to the root folder by running bash setup_folders.sh.
2. Train models by running run_all.sh training from within src/submit_scripts/. Model checkpoints will be stored to the checkpoints/<dataset> folder, and during training / analysis progress information will be saved to logs/<analysis_type>.
3. Subsequently, analyses can be conducted by running run_all.sh <analysis> from within src/submit_scripts where analysis is one of swapping | retraining | gradients | probing | centroid_analysis.
4. Individual analyses can be visualised using the corresponding notebooks (visualise_layer_swapping.ipynb, visualise_layer_retraining.ipynb, visualise_gradients.py, visualise_probing.ipynb), that start with cells for the control setup (section 3.2), followed by cells the main results analysis (section 4).
5. Centroid analysis can be performed using visualise_centroid_analysis.ipynb, after first executing visualise_mmaps.ipynb for all models / datasets, to compute the generalisation scores used in the centroid analysis correlation analysis.
6. Afterwards, summary visualisations can be computed using summarising_visualisations.ipynb.
7. For the appendix experiments using the 1.3B models, execute run_all.sh <mode> from within src/submit_scripts_big/ first using training, followed by swapping and centroid_analysis.