xjtupanda/T2Vid-Synthetic
收藏Hugging Face2024-12-02 更新2024-12-14 收录
下载链接：
https://hf-mirror.com/datasets/xjtupanda/T2Vid-Synthetic
下载链接
链接失效反馈官方服务：
资源简介：
T2Vid-Synthetic是一个从语言数据中衍生的合成“视频”指令数据集，旨在丰富视频训练语料的指令多样性。数据集从长文本QA数据集（如LongAlpaca-12k和LongQLoRA-39K）中提取，每个样本可以抽象为一个（长上下文、指令、答案）三元组，其中长上下文被分割成块并嵌入到图像中。数据集包含8750个来自LongAlpaca的样本和10076个来自LongQLoRA的样本，还提供了原始文本JSON文件以便用户自定义合成或进一步扩展。
This is a synthetic video instruction dataset derived from language data. It is designed to enrich the instruction diversity of video training corpus. The dataset is curated from long text datasets, with each sample containing a long context, instruction, and answer triplet. The context is segmented into images, and the instruction and answer are extracted from the Q&A part of the text sample. The dataset can be mixed with video training data to enrich the diversity of instruction.
提供机构：
xjtupanda
