Xander Davies, Lauro Langosco, and David Krueger
This is the repository for "Unifying Grokking and Double Descent," appearing at the 2022 NeurIPS ML Safety Workshop.
See toy_grok_dd.ipynb for the toy model results, and grok-replication/ for code replicating grokking (which can be used to replicate model-wise grokking).