ContinualConfig

Continual-learning stages over a single base run.

Each entry in datasets becomes one training stage: the base run is copied with its data replaced by that entry, trained in order, and each stage starts from the previous stage's weights (fresh optimizer). All stages run inside one experiment and are scored on all configured benchmarks.

Example: run: data: {...} # ignored; the per-stage data below is used model: {...} algorithm: {kind: sft, ...} continual: datasets:

• {dataset_name: corpus_a, transforms: [...]}
• {dataset_name: corpus_b, transforms: [...]}
• {dataset_name: corpus_c, transforms: [...]}

Attributes