Optimizing Training Data Diversity with DataSpinner: An LLM-driven Curator-Agent Approach
Year: 2026
Type: conference-paper
Abstract: Synthetic data generation has emerged as a scalable alternative to human-curated datasets for fine-tuning task-specific language models. However, uncontrolled generation pipelines risk distributional collapse, where the training corpus converges toward a narrow, low-variance distribution that undermines model generalization. In this paper, we propose a curator agent, an LLM-driven component of the DataSpinner framework that treats dataset diversity as a first-class optimization objective, and evaluate the approach on the NLIs to JSON-formatted action sequence compilation task for mobile robots. The curator agent is designed as a closed-loop framework: it analyzes the current corpus distribution across five interpretable dimensions (complexity, logic type, domain, sentence structure, and specificity) and steers a secondary augmentation agent to generate data that fills identified coverage gaps. We instantiate the framework in a single-iteration A/B configuration and compare it against a non-curated baseline of equal size. Results show that the curator-driven pipeline produces measurably more diverse training data than classical augmentation, while evaluation across five Qwen2.5Instruct model sizes (0.5B–14B) revealed that although fine-tuning improves performance over zero-shot prompting, smaller models favor tightly focused datasets while larger models can accommodate greater training diversity.
Institutions: Obuda University, Technical University of Košice
Language: English
Topic: Topic Modeling
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7204204986