I'm going to start by assuming that you don't actually want 10 protein sequences from the same species, even though you write this in your comment on M__'s (deleted) answer. This will not be a very informative sample of variation in a given protein, except for extreme examples like antibodies.
To get a good sample of variation you almost certainly want to look across species. The paper that you link to does this by using the UniRef database, they then cluster proteins at the 90% similarity threshold. That is well beyond the variation that you'd expect within a species. See their section "Training datasets":
We train our masked language models on a large corpus of protein sequences, UniRef90 (Suzek et al., 2007), retrieved in July 2018. This dataset contains 76,215,872 protein sequences filtered to 90% sequence identity.
You could use an approach like that in the paper to get similar-ish clusters of proteins whose variation you can then analyze. You can use this, but I am somewhat suspicious of that, without reading the paper in too much depth, because we know rather a lot about orthology/paralogy and how it affects function in these cases. Especially in these very elaborate not very introspectable models I worry about what they will do with inadequately curated data.
So instead I'd probably want to first identify the orthology groups (gene families) of interest. These are generally manually curated groupings of proteins.
Once you have the orthology groups that you care about, you can use a database such as KEGG or OMA to identify the specific orthologs (genes) that belong to the group. Then you will have your dataset.