Chain-of-Thought (CoT) grounding dataset
收藏资源简介:
CoT数据集是由清华大学深圳国际研究生院和阿里巴巴集团的研究人员创建的,包含90,000个样本,每个样本都标注了详细的推理链。该数据集旨在指导模型通过监督微调沿着正确的推理路径进行,从而提高其推理能力。数据集的创建过程包括使用Qwen-VL-MAX模型生成推理链,并经过人工验证确保准确性。CoT数据集的应用领域是通用视觉定位任务,旨在解决现实世界场景中涉及复杂和多模态上下文的视觉定位问题。
The CoT dataset was created by researchers from Tsinghua Shenzhen International Graduate School and Alibaba Group, consisting of 90,000 samples each annotated with a detailed reasoning chain. This dataset is designed to guide models to follow correct reasoning paths through supervised fine-tuning, thereby enhancing their reasoning capabilities. The dataset construction process includes generating reasoning chains using the Qwen-VL-MAX model, followed by manual verification to ensure accuracy. The CoT dataset targets general visual grounding tasks, aiming to address visual grounding problems involving complex and multimodal contexts in real-world scenarios.




