Giesen/TA-116k

TA-116K 数据集是 ICML 2026 论文 *Task-Aware Mechanism: Hybrid MoE Vision Tower Towards Holistic Video Understanding* 中用于 Inductor 训练和测试的数据集。该数据集主要源自 LLaVA-Video-178K，后者包含大约 2M 个 Question-Answer 对。经过去重和修复后，使用 Deepseek-R1 进行标注得到 TA-116K。论文中使用的训练数据约为 116k 样本，而本仓库提供的完整数据集约包含 1.1M 行。各文件的编号仅代表采样批次，使用时可以直接混合。TA-116K Dataset is used for training and testing Inductor in the ICML 2026 paper *Task-Aware Mechanism: Hybrid MoE Vision Tower Towards Holistic Video Understanding*. It is mainly derived from LLaVA-Video-178K, which contains approximately 2M question-answer pairs. After deduplication and correction, the data was annotated using Deepseek-R1 to obtain TA-116K. The training data in the paper contains approximately 116k samples, while this repository provides the complete dataset with approximately 1.1M rows. The numbering of each file only indicates its sampling batch, and the files can be directly mixed and used together.