v0.2.4

• QuadraticEditor is now directly importable from the main module
• QuadraticFitter.editor() now has an optional device field that you can use to create an editor on a device other than the one that the fitter was created on

v0.2.2

Various little bug fixes, mainly:

• QuadraticFitter no longer has embarrassing crash on CUDA
• We now work on PyTorch <2.0

We now also have very basic support for complex numbers, although full correctness is not yet guaranteed (#9). This is an ALPHA feature.

v0.2.1

Removes HuggingFace datasets and transformers as hard dependencies. These libraries are now only imported when the relevant functions from the scrubbing module are imported.

This release also removes a handful of vestigial utility functions which were not used by the core API.

v0.2.0

Added Oracle LEACE implementation (#2) which achieves even more surgical edits than standard LEACE when ground-truth concept labels are available at inference time. The classes OracleFitter and OracleEraser are designed to work almost exactly like LeaceFitter and LeaceEraser except that OracleEraser requires an extra z positional argument in its forward method.

This release also fixes a subtle bug in our covariance matrix shrinkage implementation which caused NaN results when the data has zero variance (#3).

v0.1.0

Refactoring

ConceptEraser has been split into two separate classes, LeaceFitter and LeaceEraser. This makes it easy to save the fitted erasure function by itself in a compact format, without also saving the covariance and cross-covariance statistics used to create it.

Algorithmic changes

We now use the asymptotically optimal shrinkage formula from this paper to shrink the covariance matrix of X toward a multiple of the identity matrix. Under weak assumptions, this provably speeds up the convergence of the covariance matrix estimate toward the population covariance matrix. Prior versions had used the raw sample covariance matrix with no shrinkage, which can cause numerical instability and very suboptimal edits when the sample size is low.