wassname's Abliterator (with baukit not transformerlens)

uv sync

then open nbs/01_abliterate.ipynb in Jupyter Notebook.

TODO:

• tidy up
• use https://github.com/wassname/activation_store to cache large numbers of examples to disc

See also

I use

• performance: wikitext perplexity

• compliance: https://huggingface.co/datasets/wassname/genies_preferences

• FailSpy's abliterator: https://github.com/FailSpy/abliterator

  • uses datasets Undi95/orthogonal-activation-steering-TOXIC vs tatsu-lab/alpaca

• a more advanced method, instead of removing all diferences, it removes the ones that are predictive, https://github.com/EleutherAI/concept-erasure https://github.com/EleutherAI/conceptual-constraints/blob/main/notebooks/concept_erasure.ipynb

  • which uses the HANS dataset:

• https://github.com/Sumandora/remove-refusals-with-transformers/

  • uses advbench/harmful_behaviors vs alpaca_cleaned

• https://huggingface.co/blog/mlabonne/abliteration