UCSC-VLAA/VLM-CapCurriculum-TextReasoning-Data

VLM-CapCurriculum-TextReasoning（D_text）是一个用于视觉语言模型后训练中第二阶段文本推理的数据集，源自ORZ-Math-13k的精选子集，包含13,181个纯文本数学问题，无图像内容。该数据集旨在巩固感知（第一阶段）和视觉推理（第三阶段）之间的文本推理能力，用于ICML 2026论文中的分阶段后训练方法。每个数据行包含问题、答案、预计算的通过率（pass_rate）等信息，通过率基于Qwen3-VL-8B-Instruct模型的16次rollout计算得出，表示样本难度，可用于能力与难度课程实验。数据以JSONL格式存储，支持按难度排序进行课程学习。VLM-CapCurriculum-TextReasoning (D_text) is Stage-2 textual-reasoning data for the staged post-training recipe in From Seeing to Thinking: Decoupling Perception and Reasoning Improves Post-Training of Vision-Language Models (ICML 2026). It is a curated ORZ-Math-13k subset consisting of challenging text-only math problems used to consolidate textual reasoning between the perception (Stage 1) and visual-reasoning (Stage 3) RLVR stages. The dataset includes 13,181 train rows, is text-only with no images, and features a precomputed pass_rate derived from 16 rollouts of Qwen3-VL-8B-Instruct, serving as an empirical difficulty signal for curriculum experiments.