1

i am trying to build question to sql like text to sql using fine tuning model

i have created example jsonl (dataset) with list of question and sql query

how can i train fine tuning model to know list of columns? because that custom model is giving wrong column name

if solution is we add list of columns to jsonl but what will prompt question look like?

i don’t want to add prompt in create completion because it will increase token of each prompt

in dataset or jsonl

{
“prompt”: “How much i spend on this month?”,
“completion”: “SELECT SUM(cost) FROM TABLE_NAME WHERE invoice.month=‘202303’”
},

like this

it doesn’t have mentioned list of columns

there are around 10-15 columns

so i want fine tuning custom model to know list of columns

how can i do that?

1 Like
2

You don’t need to fine tune for this.

3

but there will be increase token of each prompt in create completion then?

like i mentioned -

i don’t want to add prompt in create completion because it will increase token of each prompt

4

Consider using gpt-3.5-turbo as it is 10x cheaper. But if you want to fine-tune go ahead, make sure your jsonl dataset follows proper guidelines

@ruby_coder also wrote an in depth tutorial in the forum, read that as well.