Inference on Custom Use Case w/o fine tuning

[Mrs-Hudson]

Hi,
Thanks for sharing this work.

1. What's the best way to perform a quick analysis on a custom domain w/o fine tuning (zero shot or few shot)?
   How should I prepare my input dataset and prompts?
2. Is it possible to perform inference on the model itself without the retriever components etc (my schema is already small) .
3. Is it possible to use vLLM for inference?

A notebook showing datatset prep and inference for a custom use case would be super helpful!

[lihaoyang-ruc]

1. To quickly analyze a custom domain database, the optimal approach is to utilize our fine-tuned models for zero-shot inference. We have created a demonstration available at https://github.com/RUCKBReasoning/text2sql-demo, which provides guidance on leveraging your databases for inference. This repository is designed to assist individuals aiming to replicate the results outlined in our paper.
2. Certainly, you can modify the specified lines to integrate support for your schema filter: https://github.com/RUCKBReasoning/text2sql-demo/blob/1f0d7b4a53a863f00434928f770053052b0615a9/text2sql.py#L139C9-L141C90
3. Yes, CodeS is built upon StarCoder, which is supported by vLLM.