"Compositional Reasoning and Preference Optimization Dataset"

"This dataset is a large-scale, multi-stage dialogue and preference alignment dataset designed to enhance the reasoning capabilities of Vision-Language Models (VLMs) in computational aesthetics and image cropping. Built upon the public CADB and GAICD databases, it deconstructs the subjective aesthetic evaluation process into an \"Analysis-Proposal-Decision\" workflow.The dataset consists of two core subsets:Compositional Reasoning Subset: A comprehensive collection of 13,204 dialogue samples that integrates visual reasoning with aesthetic judgment. It first provides structured JSON annotations (9,204 samples) to explicitly identify core compositional elements (e.g., rule of thirds, leading lines) and generate rule-based candidate crops. Building upon this context, it incorporates 4,000 high-quality multiple-choice samples guided by Mean Opinion Scores (MOS), providing ground-truth annotations for selecting the most aesthetically pleasing crop from the generated candidates.Preference Optimization Subset: 8,000 preference pairs tailored for Direct Preference Optimization (DPO). It includes chosen and rejected responses, deliberately incorporating \"hard negatives\" (suboptimal crops with minor compositional flaws) to improve the model's ability to evaluate fine-grained aesthetic trade-offs."