# DataFlow: An LLM-Driven Framework for Unified Data Preparation and Workflow Automation in the Era of Data-Centric AI

**Authors:** Hao Liang, Xiaochen Ma, Zhou Liu, Zhen Hao Wong, Zhengyang Zhao, Zimo Meng, Runming He, Chengyu Shen, Qifeng Cai, Zhaoyang Han, Meiyi Qiang, Yalin Feng, Tianyi Bai, Zewei Pan, Ziyi Guo, Yizhen Jiang, Jingwen Deng, Qijie You, Peichao Lai, Tianyu Guo  
**arXiv:** [2512.16676](https://arxiv.org/abs/2512.16676)  
**Date:** 2025-12-18  
**Summary page:** https://www.summarizepaper.com/en/paper/2512.16676v1/

## Key points

- The DataFlow Technical Report presents a comprehensive and innovative framework for data preparation in Large Language Models (LLMs)
- DataFlow addresses the growing demand for high-quality data in LLMs and challenges posed by current ad-hoc practices
- DataFlow provides a unified and extensible solution with system-level abstractions enabling modular, reusable, and composable data transformations
- Designed with a PyTorch-like programming interface for usability, extensibility, and maintainability
- Offers nearly 200 reusable operators and six domain-general pipelines spanning various domains such as text, mathematical reasoning, code, Text-to-SQL, agentic RAG, and large-scale knowledge extraction
- DataFlow-Agent automates translation of natural-language specifications into executable pipelines through operator synthesis and iterative verification
- Extensive experiments demonstrate effectiveness of the framework across diverse data preparation scenarios
- Resulting datasets consistently match or surpass state-of-the-art baselines in quality and performance
- Mathematical reasoning data outperforms high-quality synthetic baselines on various benchmarks
- Pipelines achieve significant improvements in execution accuracy compared to existing corpora while using fewer training examples
- Training on only 10K samples can yield substantial gains in data efficiency compared to models trained on larger datasets when combining text and code data into a unified corpus with DataFlow.

## Summary

The DataFlow Technical Report presents a comprehensive and innovative framework for data preparation in Large Language Models (LLMs). The report addresses the growing demand for high-quality data in LLMs and the challenges posed by current ad-hoc practices. To overcome these challenges, DataFlow provides a unified and extensible solution with system-level abstractions that enable modular, reusable, and composable data transformations. Designed with a PyTorch-like programming interface, DataFlow ensures usability, extensibility, and maintainability. The framework offers nearly 200 reusable operators and six domain-general pipelines spanning various domains such as text, mathematical reasoning, code, Text-to-SQL, agentic RAG, and large-scale knowledge extraction. In addition to this rich ecosystem of operators and pipelines, DataFlow-Agent automates the translation of natural-language specifications into executable pipelines through operator synthesis and iterative verification. Extensive experiments on six DataFlow-implemented pipelines demonstrate the effectiveness of the framework across diverse data preparation scenarios. The resulting datasets consistently match or surpass state-of-the-art baselines in terms of quality and performance. For instance, mathematical reasoning data outperforms high-quality synthetic baselines on various benchmarks. Similarly, pipelines achieve significant improvements in execution accuracy compared to existing corpora while using fewer training examples. By combining text,and code data into a unified corpus (), DataFlow shows that training on only 10K samples can yield substantial gains in data efficiency compared to models trained on larger datasets. This highlights the framework's ability to produce high-quality supervision across diverse domains. Overall, emerges as an end-to-end system for LLM-based data preparation that offers a rich ecosystem of operators and pipelines. With its focus on principled abstractions and scalability, DataFlow enhances programmability, reproducibility, and data quality in LLM workflows. It serves as a foundational tool for building semantically rich and scalable data preparation pipelines that improve performance across various domains.