AutoGeo-100K
AutoGeo-100K is the largest dataset of geometry shapes, comprising 100,000 images of geometry shapes along with their corresponding natural language descriptions and clauses that determine the semantic meaning of the shapes. This dataset is generated automatically by combining 77 clauses through the AutoGeo pipeline, which includes a Rule-based Clause Selector, an image generation module capable of producing various images with the same semantic meaning, as well as a language generation module that generates 20 caption templates for each clause. We propose Geometry Captioning Tasks (GC) based on the AutoGeo-100K dataset and evaluate the performance of several state-of-the-art MLLMs on this task. Additionally, we compare the performance differences of baseline models and models tuned on smaller subsets of the AutoGeo dataset (AutoGeo-10K, AutoGeo-30K, AutoGeo-50K) on the Geometry Question and Answer Task (GQA). Our findings indicate a performance increase in both tasks as the volume of tuning data from the AutoGeo dataset increases, highlighting the importance of having a large amount of geometry data for training and evaluation purposes. Given the limited data volume of current geometry datasets and the challenges associated with obtaining geometry image and caption data, our goal is to provide the research community with a comprehensive dataset for studying MLLMs in the context of geometry. Additionally, we present a methodology for generating a large quantity of new geometry data, which we believe will be beneficial for further research and development in this field.
