This is an implementation of the compilation from C-RASP programs into transformer encoders, as specified in the paper. More details to come... :)

Implement Bools

• And
• OR
• Comparison...

Implement Counts

• Conditional
• Addition
• Subtraction
• Min
• Max
• Constants
• Design choices: Hide BOS. Also need another FFN to ensure the BOS count is always 0, otherwise it screws with the rest

Utilities

• When pretty-printing, mask out the counts that have been overwritten by LayerNorm during Comparison
• Make a more user-friendly interface for writing programs - a parser?
• Make it very streamlined to use as language acceptor/rejector
• Companion program to just simulate C-RASP programs

Language Modeling

• I said we could define simple language models using C-RASP, so we should test it out
• Presumably this can be used to edit existing language models. For instance, we can force it to never generate X if it has not seen at least 3 Y's. This can be done by adding fresh dimensions and inserting a C-RASP program which checks whether or not there have been 3 Y's or not. Then we can add a huge coefficient to this Boolean value and add it to the dimension which stores the probability of X being the next token. We'd also have to muck around with the final layer, but depending on how the language model is defined this is maybe ok.

More

• How much of this can be applied to S-RASP?
• Experiments -- initialize a C-RASP transformer and fine-tune on something