For any future gemmascope releases, could you add ids and/or hashes?
#10
by
lodrick-the-lafted
- opened
I was revisiting old experimental code which loads a gemmascope SAE (in the hopes of interpreting control vectors I've generated) and I am loading a file named params.npz which I downloaded a number of months ago. Yes, it's absolutely my own fault for failing to keep track of which model variant I am using, but it would be helpful for future releases to include an id in the actual filename, or some method of distinction, even if it's just list of hashes to corresponding models.