Overview

DataChef-32B is a specialized 32 billion parameter large language model developed by yichengchen24, designed to automate the creation of data processing pipelines, referred to as "data recipes." Its primary function is to facilitate the adaptation of base LLMs by generating optimal training corpora from raw data sources, tailored for specific benchmarks. The model was trained using online reinforcement learning, employing a proxy reward system to predict and maximize downstream performance of adapted LLMs.

Key Capabilities

• Automated Data Recipe Generation: DataChef-32B outputs complete, executable data processing pipelines given a target benchmark and available data sources.
• LLM Adaptation Optimization: It streamlines the process of preparing high-quality training data, which is crucial for fine-tuning LLMs for specific tasks.
• Performance Comparable to Human Experts: The model generates practical recipes that achieve performance levels similar to those curated by human experts.
• Demonstrated Performance Improvement: A recipe generated by DataChef-32B for Qwen3-1.7B-Base achieved a score of 66.7 on AIME'25, outperforming the standard Qwen3-1.7B in the math domain.

Use Cases

• Efficient LLM Fine-tuning: Developers can use DataChef-32B to quickly generate optimized datasets for fine-tuning their LLMs for various benchmarks.
• Reducing Manual Data Engineering: It significantly reduces the labor-intensive process of manually designing and implementing data processing pipelines for LLM training.
• Targeted Model Improvement: Ideal for adapting LLMs to specific domains or tasks where high-quality, domain-specific data is critical for performance.