$ the-wire · showcase
Slow PGD ladders read out from one attack run
By RepoJournal · Filed · About Interpretability · Composed from the cited sources · methodology
param-decomp now reads a full ladder of slow PGD checkpoints from a single attack, logging each step count as its own metric instead of requiring a rerun per read-out point.
Read slow PGD ladders from one attack goodfire-ai/param-decomp
by PD User (shared)
SlowPGDReconLoss is evaluated once per step count in a strictly ascending read_out_steps list and logged as SlowPGDReconLoss_<n>step, so a single attack now yields the whole sweep rather than one attack per checkpoint. Fast probes continue to land under PGDReconLoss_<n_steps>step, each eval metric type appears at most once, and the maintained language-model configurations already carry the new ...
