MAGIC: Generating Self-Correction Guideline for In-Context Text-to-SQL is a dataset released by Microsoft in 2024 for generating self-correction guidelines in text-to-SQL tasks. The related research paper is "MAGIC: Generating Self-Correction Guideline for In-Context Text-to-SQL", which aims to correct erroneous SQL queries through agent interaction and generate corresponding correction guidance rules.

This dataset is built on the BIRD dataset, which contains 12,751 question-answer-SQL pairs extracted from 95 large databases, covering 37 professional domains such as blockchain and healthcare, and introduces external knowledge to increase the complexity of query generation. The dataset primarily records the interaction trajectories of agents correcting SQL queries during the iterative process, helping to study error correction mechanisms in text-to-SQL tasks.

Dataset Composition

Each trajectory data file contains the following fields:

• prompt: The prompt input to the agent
• response: The agent's response to the prompt
• caller_agent: Identifies the agent invoked in each iteration, with values including feedback_agent_call_, correction_agent_call_, and manager_revises_prompt_iteration_

Citation

@article{askari2024magic,
  title={MAGIC: Generating Self-Correction Guideline for In-Context Text-to-SQL},
  author={Askari, Arian and Poelitz, Christian and Tang, Xinye},
  journal={arXiv preprint arXiv:2406.12692},
  year={2024}
}