[kodra]

I found it interesting (and kinda worrisome) that some fine-tuned Llama 2 models are trained with GPT-4 generated data. So if GPT-4 got something wrong, the error would propagate to a lot open source models.