8 likes
·
13.0K reads
9 comments
Hello sir, I am intrigued by your article. I have got much more understanding, however I wanted to understand that is task specific to a particular schema?
I see that you are not passing the fields in the prompt, so how is it picking up about which field to use
In this case, yes, I have fine-tuned the model for a specific schema. You can try using your own schema.
As for your second question, I'm not sure I understand it fully. If you're asking how the model predicts which fields to use based on the prompt, it has an overall understanding of the schema, including the tables and their columns. I prepared the training dataset in a way that includes all the information about our schema, which gives the model a better understanding of our task
Rupesh Gelal thank you so much for your reply.
Yes your second answer is in inline with my query. Also is it possible to create a universal text to sql query, such that user just has to define schema
Meet Vora Yes, those systems are called Natural Language Interface to Databases (NLIDB). You can look them online
can you plz provide me this csv file ?Rupesh Gelal
What would u suggest - fine tuning on custom data set vs using vectordb for passing required metadata abt the tables
Its very difficult to fine tune 1000’s of table across multiple database etc…
We noticed Hallucination of llm cannot be reduced using fine tuning
Do u have code for transfer learning on text to sql for gpt3 - so that it does its best in sql generation
Thanks for this helpful article about fine-tuning GPT-3 for natural language to SQL conversion. For those who are interested in using ChatGPT, I recommend reading this article (blog.devart.com/how-to-use-chatgpt-to-writ…) about how to use ChatGPT to write SQL JOIN queries
Thanks for this help article. I want to know how big is the dataset, and how many tables it can cover after fine tuning.
Its really good, it would be great if the author can share the training file that he has used to train the model.