- Probeless method implementation has been added in the
interpretationmodule - Data balancing and binary annotation methods are now part of the
datamodule - New options for transformers extraction: Decompose layer activations in separate files, Save activations only for certain layers
- Dedicated Activations Writer class that abstracts away representation saving from representation extraction
- Fixed Packaging error that included debug code
- Linear regression probe did not supply the correct arguments (Issue #18)
- Transformers extractor can now optionally take a tokenizer name instead of just a model name