yunfengwang/TVP-Training-Data

该数据集是用于Thinking with Visual Primitives（TVP）PyTorch实现的多阶段训练数据。它包含预训练和SFT（监督微调）两部分：预训练数据（146K样本）基于COCO数据集，提供物体定位（标签和边界框）标注；SFT数据包括物体定位（30K样本，含结构化思维链和15%负样本）、计数（8K样本，含边界框定位和思维链）、空间推理（3K样本，CLEVR风格）、迷宫导航（5K样本，使用点基元进行程序化生成）和路径追踪（3K样本，点序列）。数据格式为JSONL，坐标归一化为[0, 999]的整数。数据集支持视觉推理、物体定位、计数、空间推理、迷宫导航和路径追踪等任务，并包含视觉基元（如边界框和点序列）的标注。图像来源包括COCO 2017数据集（用于定位和计数任务）和程序化生成的图像（用于迷宫和路径任务）。This dataset contains all training data for the multi-stage TVP (Thinking with Visual Primitives) PyTorch implementation. It includes pretrain and SFT (Supervised Fine-Tuning) splits: pretrain data (146K samples) is COCO-based grounding (label + bounding box) annotations; SFT data comprises grounding (30K samples, with structured thinking chain and 15% negatives), counting (8K samples, with bounding box grounding in chain-of-thought), spatial reasoning (3K samples, CLEVR-style), maze navigation (5K samples, using point primitives procedurally generated), and path tracing (3K samples, point sequences). Data format is JSONL with coordinates normalized as integers in [0, 999]. The dataset supports tasks such as visual reasoning, grounding, counting, spatial reasoning, maze navigation, and path tracing, and includes annotations for visual primitives (e.g., bounding boxes and point sequences). Image sources include COCO 2017 dataset (for grounding and counting tasks) and procedurally generated images (for maze and path tasks).