Video-Reason/VBVR-MultiStep-Bench
收藏资源简介:
VBVR-MultiStep-Bench是VBVR-MultiStep训练语料库的公开评估分割部分,设计用于长期多步图像到视频(I2V)推理评估。该数据集包含36个参数化任务,共180个实例,每个实例包含5个文件:first_frame.png(模型调节图像)、prompt.txt(自然语言任务合同)、final_frame.png(目标端点)、ground_truth.mp4(参考展开)和question_metadata.json(问题元数据)。数据集主要用于I2V系统在固定五工件合同下的轨迹级评估,支持盲人两两比较评估。数据集是完全合成的,不包含第三方受版权保护的内容。
The frozen 180-instance public evaluation split released alongside the VBVR-MultiStep training corpus. Designed for long-horizon multi-step image-to-video (I2V) reasoning evaluation. This dataset contains 36 parameterized tasks with 180 instances (5 per task × 36). Each instance includes five artifacts: first_frame.png (model conditioning image), prompt.txt (natural-language task contract), final_frame.png (target endpoint), ground_truth.mp4 (reference rollout), and question_metadata.json (question metadata). The primary use is trajectory-level evaluation of I2V systems under a fixed five-artifact contract, supporting blind human pairwise judging. The dataset is fully synthetic with no third-party copyrighted content embedded.




