Caching

Language model activations

• encoding.utils.ActivationCache persists multi-layer activations to a single .pkl per config

• encoding.utils.LazyLayerCache loads metadata fast and layers on demand

• Cache keys include story, model name, context type, lookback, dataset_type, etc.

Speech activations

• encoding.utils.SpeechActivationCache and SpeechLazyLayerCache mirror the API for speech features with optional time arrays

Model artifacts

• encoding.utils.ModelSaver records hyperparameters and metrics per run directory (hash + timestamp)

• Use list_runs() to enumerate previous runs