Convert common context information to all prompts in fine-tuning from Python

API
1

Hi,
I’m using Chatgpt3 to train it on SQL queries.
For now I’m using the “Playground” section where I declare the shape of the db + promp&queries for fine-tuning. Now I’d like to do fine-tune from Python, but the db structure definition isn’t really an example prompt=>response, but it is precisely common to all the examples.
I specify that I need to use Python because I have exceeded the maximum context length of 4097 tokens (if anyone knows how to remove this constraint from Playground, yo are welcome).
How do I communicate the database structure common to all prompts from Python to develop a new specific model?
The most inefficient alternative of all is to repeat the context structure in all prompts, but it is very inefficient and expensive.

2

What have you tried so far?

Can you show some code? Your prompt? Details help…

Thanks!

3

Hi @PaulBellow, i have only the prompts from Playground (i have no code for Python):
I’ll give you an example:
PART COMMON TO ALL EXAMPLES (context)
Write a query in SQL Server from natural language on the following database:

  • Customer Table (columns: customer name, geographical area, …)
  • Order Table (columns: order date, invoice date, quantity …)
  • List item
    …
    SPECIFIC PART with Query examples (real fine-tuning):
  1. What are the customers? => select customer name from customer table
  2. What are the quantities ordered => select quantity from Orders Table
    …
4

What settings are you using? Temperature, frequency_penalty, etc.

How are the outputs not what you expected?

Have you searched the forums? I think a few others have tried to do what you’re attempting… I think…