Why did you chose to train in two steps?

[liang8qi]

Thanks for your sharing, your attempt is very interesting and valuable.

However, I have some questions about the training process.

I notice that ChatDoctor is first trained using 52K instruction-following data from provided by Stanford Alpaca, and then finetuned on the your specific data.

Why not finetune the model using a mixture of two parts of the data?

What is the insight of this finetuning process? What is the insight behind this finetuning model?

Have you ever tried to train with two pieces of data mixed together?