Synthetic dataset generation - Brief howto built on top of llama.cpp

[paschembri]

I’am currently exploring finetuning smaller models for classification / fill-mask.

I wrote a quick intro on how to produce synthetic dataset in a few lines of python.

It covers:

• the raw generation
• the automatic labelling/classification

If you have performance tips for this kind of tasks I’m all ears 👀. Now I have to look on how to implement distilbert using ggml 😱

[ianscrivener]

Suggest you first explore grammer functionality in llama.cpp and the excellent grammer builder helper app. I suspect you'll find them valuable.

[paschembri]

I did and it's great - I'm currently looking on how to backport the feature to the ggml lib (#2494)