Preliminary results here: https://docs.google.com/document/d/1LCjjnUPN51gHl_rmCWEmmtbY-Wu1dixzOif14e-7i-U

Structure of the repository:

• generate many hidden states with generate_main.py (it uses Collin's code from the zip files, copied in utils_generation)
• find direction using CCS recursively by adding new constraints each time using generate_ccs_dirs_main.py (it uses Collin's code from the github repo + bugfixes + custom things I want for recursion & debuging, the core functions are in utils.py). You can see what I use it for by looking at script.sh (it's a bit messy).
• analyze the directions found by CCS by running analyze_ccs_dirs_main.py (looking at a single layer) or analyze_ccs_dirs_layers_main.py (looking at multiple layers at a time) as Python notebooks.
• get a sense of what the CCS loss does by running vizualisation.py as a Python notebook.
• see what happens when you mix directions by running mixing_directions.py as a Python notebook.

Note: evaluate.py, generate_ccs_dirs.py, evaluate.py, and analyze_ccs_dirs.py legacy files from when I only looked at datasets composed of a single prompt. They are not used anymore.