If you train LLMs on code from GitHub it will output what the average GitHub repo has in it. That is, "competent level", not expert or virtuoso.

And it will also carry over bugs found at the "competent" level.