Hi,
I just noticed you mentioning anagram manipulation issues, this is often down to the way the model tokenizes word parts, to you ENTJ looks like 4 letters, but to the model it looks like this 
2 tokens.
As a test, can you try inserting a separator character such as - (dash) to give this E-N-T-J, and the model will now see that as 
4 separate tokens.
You can then extract the - to rebuild your string. See how that works for you.