open galleon
Hey people I have a problem with the SQL translate example.
Does anyone know any technique for e.g: I don't need to pass the schema everytime. It's just the tokens are too much when there's 100s of tables.
If fine tuning is the way , then how can I make the model memorize my table schema ? I've tried but it doesn't work because its not how we are supposed to train right ?. Then what is the perfect way to let the model know my table schema and business process ?
fading river
What are you hoping to accomplish? If it's structured data that maybe you want to retrieve with regular english why not use embeddings?
open galleon
@fading river . I'm trying to convert natural language to SQL queries which davinci-003 or codex does perfectly. But there's a twist. It requires me to pass the table schema everytime in the request which is not possible when I have 100s of tables right ?. Also what do you mean by embeddings can you explain ?
fading river
In your case it may be best to separate each table and only provide what's needed programmatically .Do you explicitly call for the table? What would an example prompt look like?
You could also fine tune it. Feed it some examples of good prompts, try some that failed, keep going. It'd be much easier than embeddings
Try some as in prompt and see if it passes this time
open galleon
Here is an example request but it realtime there would be more tables with huge number of columns
SELECT
fading river
Ok well I guess you could go two ways
If you are always going to call for the table. That's pretty easy. You can just programmatically retrieve its reference
open galleon
Yeah I've tried fine tuning with some examples. but I cannot get it to memorize the column names it always gets a random column name or gets the query wrong. it is worse compared to davinci-003
fading river
Yeah I've noticed the same. Ironically it's not very good with numbers
The example you showed me seems ideal though. You can easily accomplish that with a simple search function in any programming language
Could probably even do it in a spreadsheet formula
open galleon
If you are always going to call for the table. That's pretty easy. You can just ...
so you mean I need to programmatically Identify the context of the question and tables that are required ?
fading river
Just the table
open galleon
Ok I will try that. Thanks for your time and help
fading river
No problem. you may not even need to fine tune if you're saying it already works well with the tables you provide
open galleon
No problem. you may not even need to fine tune if you're saying it already works...
Yes. But sometimes it doesn't because it does not know some of the complex business process that we have set. Maybe that can be fixed by fine tuning but still that model is not as good as davinci-003
fading river
In that case it'd be worth fixing the bad answers and refeeding them. Unless some prompt magic can do the trick
open galleon
yeah I will try that . I also hope there is some prompt magic 
unkempt scroll
Have you tried Babbage, or another code completion model?
open galleon
@unkempt scroll yes but other than davinci -003 or codex, most of them doesn't get the query correctly and accurately
unkempt scroll
Yeah I've tried fine tuning with some examples. but I cannot get it to memorize ...
Your query should be something like:
Using SQLQuery, write a query that gets the value I need, from the columns named 'column1,column2,column3'.
The table is called 'myTable' and the user is called 'myUser'
Or get gpt to fix your table columns and import your data to a new table
open galleon
Your query should be something like:
Using SQLQuery, write a query that gets th...
So you're saying that my training examples should be like this ?
fading river
So you're saying that my training examples should be like this ?
Why not try using a small amount of data to benchmark? You could try 20 prompts using different styles and compare the results