Added attention pattern and logit lens analyses for corrupt cache in both activation patching and path patching.
Had to move st.extenders for attention patterns and logit lens to the app.py level because you can't have extenders within extenders, which would have happened with activation patching -> logit lens (as an example) otherwise. This does not affect the behaviour of the UI that I can detect.
The big changes in dynamic analyses can be largely ignored - those are just removing the st.extender items from it, and altering the indentation level of everything that's left, plus adding some keys.
Tested:
Made sure that the analyses from clean and corrupted were different.
Made sure no Streamlit errors were caused by having the clean analysis, corrupt activation analysis, and corrupt path analysis all open at once in Streamlit.
Performed:
Added attention pattern and logit lens analyses for corrupt cache in both activation patching and path patching.
Had to move st.extenders for attention patterns and logit lens to the app.py level because you can't have extenders within extenders, which would have happened with activation patching -> logit lens (as an example) otherwise. This does not affect the behaviour of the UI that I can detect.
The big changes in dynamic analyses can be largely ignored - those are just removing the st.extender items from it, and altering the indentation level of everything that's left, plus adding some keys.
Tested: