compression-progress

Compression progress measures learning by computing the derivative of a compressor's performance over time—specifically, the reduction in bits needed to encode data as the model improves. It generates an intrinsic reward signal r(t) = |C(t-1)(data)| − |C(t)(data)|, returning a positive reward when compression efficiency increases, driving curiosity and exploration in learning systems.

Yes. Compression progress identifies tasks with high expected improvement, enabling curriculum generation where the system prioritizes learning experiences that maximize compression gains. This approach drives task selection and exploration strategies by rewarding observable progress in model compression.

Compression progress accepts model checkpoints (before and after states) and data files as inputs. The Python interface computes old_bits and new_bits from compress(data) calls, making it compatible with any compressor and data domain where you can measure bit reduction between sequential model states.

Use the command-line interface: `compression-progress before.model after.model data/`. The Skill computes compression metrics on your data across both model versions and returns the intrinsic reward signal based on bit reduction, enabling direct integration into learning loops and curriculum design.

Compression progress grounds intrinsic motivation in measurable model improvement rather than novelty or prediction error. It directly quantifies learning efficiency by tracking bits saved, making reward signals interpretable, reproducible, and aligned with actual compression gains in the learned representation.

Avoid compression progress when your data domain has minimal compressibility, when model improvements don't correlate with bit reduction, or when you need rewards independent of model state changes. It's most effective for domains where compression efficiency meaningfully reflects learning progress.