Do you think it helps with the systems reasoning? Are there many tla+ specs + real code bases for the llm to learn from?

Raft is a good example. And it seems to be in most models training data.