DataChef: Cooking Up Optimal Data Recipes for LLM Adaptation via Reinforcement Learning

Dimensions overall score 8.0

GitHub Code Pulse

No public code linked for this paper yet.

• DataChef automates end-to-end data recipe generation for LLM adaptation.method 95%
• DataChef-32B uses online reinforcement learning with a proxy reward to predict downstream performance for candidate recipes.method 95%
• DataChef-32B produces practical data recipes that achieve comparable downstream performance to human-expert curated recipes across six held-out tasks.result 90%
• A recipe generated by DataChef-32B adapted Qwen3-1.7B-Base to the math domain, achieving 66.7 on AIME'25 and surpassing the base model's performance.result 95%
• The success of generated pipelines is heavily dependent on the quality of the reward function in the reinforcement learning model.limitation 90%
• DataChef can potentially disrupt existing data engineering practices by offering an automated solution for data preparation and model training.market 85%
• DataChef was evaluated by comparing LLM performance from generated recipes against human-expert curated pipelines.method 90%

Competitive landscape

Competitor map is still being generated for this paper. Enable generation or check back soon.

Startup potential card