Researchers are advancing the optimization of data preparation for large language models (LLMs) with a focus on high-quality training data. A team from Fudan University and Shanghai AI Laboratory has developed DataChef-32B, an automated system that creates ‘data recipes’—pipelines that transform raw data into effective training datasets. Utilizing reinforcement learning, DataChef-32B generates comprehensive data recipes tailored to specific tasks and data sources. This innovative approach achieved notable performance, surpassing human-crafted recipes and outpacing the Qwen3-1.7B model on the AIME’25 benchmark with a score of 66.7. Key features include a Data Verifier for assessing data quality without full model training, and an integrated Code Interpreter for executing Python scripts. By addressing data assembly challenges, DataChef-32B enhances LLM capabilities and sets the stage for self-evolving AI systems. This research promises to accelerate LLM development and expands its framework to other domains requiring effective data curation, ensuring a cost-effective solution for diverse tasks.
Share
Read more