They're probably generating based on GitHub code. If I were training a code model I'd take a snippet of code, have the existing LLM explain it. Then use the ex

They're probably generating based on GitHub code. If I were training a code model I'd take a snippet of code, have the existing LLM explain it. Then use the explanation and the snippet for the test data.