WenqingCao/finevisionmax-strict-ans-ablation

FineVisionMax — Strict Numerical Ablation 是一个用于视觉语言模型中近似数字系统（ANS）表示消融研究的过滤子集。它基于HuggingFaceM4/FineVisionMax数据集，通过严格过滤规则筛选，移除任何用户或助手回合中包含15个类别（如数字、数字词、计数动词、比较词、序数词等）匹配的行。这些类别覆盖了REMOVE类（如数字和数字词）和EXPERIMENT类（如模糊量词、缺失量词、空间密度等）。数据集旨在研究ANS在视觉语言模型中的涌现，输入样本总数为24,243,243，保留2,328,928个样本，排除率为90.39%。数据集以Parquet格式提供，可通过Hugging Face datasets库加载。FineVisionMax — Strict Numerical Ablation is a filtered subset for an ablation study on the emergence of approximate-number-system (ANS) representations in vision-language models. It is based on the HuggingFaceM4/FineVisionMax dataset and applies a strict filter that removes rows where any user or assistant turn contains a match from 15 categories spanning the REMOVE class (e.g., digits, number words, counting verbs, comparisons, ordinals) and the EXPERIMENT class (e.g., vague quantifiers, absence quantifiers, spatial density, temporal counting, collective phrases). The dataset is designed to investigate ANS emergence, with 24,243,243 input samples processed, 2,328,928 kept, and an exclusion rate of 90.39%. It is provided in Parquet format and can be loaded via the Hugging Face datasets library.