DataChef-32B

HF Models | HF Demo | Paper | GitHub

DataChef-32B is a specialized large language model designed for automated data recipe generation. It was introduced in the paper DataChef: Cooking Up Optimal Data Recipes for LLM Adaptation via Reinforcement Learning.

DataChef-32B facilitates LLM adaptation by generating executable data processing pipelines (data recipes) that transform raw data sources into high-quality training corpora targeted at specific benchmarks.

Model Description

DataChef-32B addresses the manual, labor-intensive process of designing data processing pipelines. It was trained using online reinforcement learning with a proxy reward system that predicts downstream performance for candidate recipes. Given a target benchmark and available data sources, the model outputs a complete data recipe to adapt a base LLM.

Performance Highlights

Across diverse tasks, DataChef-32B produces practical recipes that reach performance comparable to those curated by human experts. Notably, a recipe generated by DataChef-32B was used to adapt Qwen3-1.7B-Base to the math domain, achieving a score of 66.7 on AIME'25, surpassing the performance of the standard Qwen3-1.7B.

Installation

To use the DataChef framework for generating your own data recipes, follow the installation steps from the GitHub repository:

conda create -n datachef python=3.12
conda activate datachef
pip install -e .

Citation

If you find this work helpful, please consider citing:

@article{chen2026datachef,
  title={DataChef: Cooking Up Optimal Data Recipes for LLM Adaptation via Reinforcement Learning},
  author={Chen, Yicheng and Ma, Zerun and Xie, Xinchen and Li, Yining and Chen, Kai},
  journal={arXiv preprint arXiv:2602.11089},
  year={2026}
}