the resolution of the input series tokenization is insufficient

[oops343]

My data have relatively large numbers(4000+), but difference between each number is actually really small (ranging from 1 to 10 maybe).
As shown in this image, the ids after meanscale tokenization are even the same for several points, which means the resolution of the bins are not enough to properly seperate them.
I wonder if there is similar situation in some of the pretraining data, and how to solve this if it is A problem.

[oops343]

i'm thinking about minmax normalization before tokenization, but not sure about how this will affect the data distribution...
the idea is to stretch the points further from each other, that makes the 'resolution' better