Uncovering the Impact of Chain-of-Thought Reasoning for Direct Preference Optimization: Lessons from Text-to-SQL

This repository is the official implementation of the paper, providing a step-by-step tutorial to reproduct the main experiment for a 7b-scale base model with 4xA100 GPUs (~1.5d).

Synthetic CoT data (Sec 3.2) and preprocessed database prompt for Bird dataset (Appendix E) can be downloaded via Google Drive for reproduction.

Code is coming in a while... (not too soon)