Hey Tiago! It was very simple actually: I split each document by paragraph break into alternating "user" and "assistant" to create a kind of call and response structure. Then I bundled all documents, with their user/assistant splits into a single jsonl file.
This would mean that if I wrote an entry in the first person as "user" it would return the next paragraph (to the user entry) as "assistant", as the model fine tunes on this call and response structure to the paragraphs. It also seems to mimic my overall style and voice. It's not perfect, of course, but was surprisingly effective!
Ha aH Guffaw weLl diD it WriTe thiS darn hEre substacK?
Hi Johan! Curious to know the details on how you converted your notes into a format that could give a good result in a fine-tunning process?
Hey Tiago! It was very simple actually: I split each document by paragraph break into alternating "user" and "assistant" to create a kind of call and response structure. Then I bundled all documents, with their user/assistant splits into a single jsonl file.
This would mean that if I wrote an entry in the first person as "user" it would return the next paragraph (to the user entry) as "assistant", as the model fine tunes on this call and response structure to the paragraphs. It also seems to mimic my overall style and voice. It's not perfect, of course, but was surprisingly effective!
Awesome! Thanks for clarifying... I will try it out! Thanks :-)
lemme know how it went for u ima do same