The script for synthesizing `fill50k`

[ernestchu]

Hi, your work is fantastic. I would like to work on some topic extending your work. The first toy problem I'm trying to solve can leverage your synthetic dataset fill50k. I might guess the script for generating the dataset is relatively simple, but I want to follow your convention and avoid possible distribution shift. Can you share the script you used?

[gabegrand]

I would also be curious to see the generation script for fill50k please!