features-across-time

Understanding how features learned by neural networks evolve throughout training

NOTE: Apologies for the sloppy code in many places. We plan to clean it up in the next few weeks.

Reproduce Pythia 410M Experiments

Run the following Python scripts from the command line to collect and plot data over training steps:

python scripts/divergences.py
python scripts/ngram_samples.py
python scripts/shuffled_samples.py

To plot existing data, use:

python scripts/plot_steps.py [--divergences] [--ngram_samples] [--shuffled_samples]