I'm surprised that this is using plain llama3.1 rather than a fine-tune. Have you checked the accuracy of the results on the common benchmarks? Also, given it p

I'm surprised that this is using plain llama3.1 rather than a fine-tune. Have you checked the accuracy of the results on the common benchmarks? Also, given it provides just the answers just based on the up/down lists, (or did I miss something?) isn't that something that could be extracted into a more efficient lookup with only a 2d grid of weights? (Or 3d if we there are group-of-genes effects)