some questions

I tried to use your code to analyze the output of a t5 model trained on Atomic. However my results are very different from yours. I use Huggingface t5 models however.

For angular change, only q has a large angular change and others are minor. Also for L1 change in decoder V and O and Ox have larger values.

Next question
I your code why you don't divide Wo and Wi matrixes on `amout_bigger`. In general what is the reason for these divisions on change and the constants for it

```
if name_of_matrix == "wi" or name_of_matrix == "wo":
change = amount_bigger * 1.048 * 1000000
else:
change = 1.048 * 1000000
```