Reproduce “CoddLLM: Empowering Large Language Models for Data Analytics” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jiani Zhang, Hengrui Zhang, Rishav Chakravarti, Yiqun Hu, Patrick Ng, Asterios Katsifodimos, Huzefa Rangwala, George Karypis and 1 more
CoddLLM introduces a foundation model specifically designed for data analytics, post-trained with a comprehensive "data recipe" to deeply understand data management concepts and multi-table relationships. The model achieves superior performance across diverse analytics tasks, outperforming its base model by 39.3% and exceeding GPT-4o's average accuracy by 4.0%.
Key insights
- Training on instruction-response pairs is significantly more effective for domain adaptation than continual pre-training on raw plain text, which can degrade instruction-following capabilities.
- Infusing broad analytics-specific knowledge (Chapter 1) provides notable gains in domain tasks, such as Table Selection and Text-to-SQL, even without explicit task-specific examples.
- Specialized text and table alignment tasks (Text-to-Schema, Row-to-Text) are crucial for enhancing complex data interpretation, retrieval, and multi-table reasoning, particularly in long-context scenarios.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.