Add dataset card for MMHU benchmark

README.md

@@ -0,0 +1,35 @@
+---
+task_categories:
+- video-text-to-text
+tags:
+- human-behavior
+- autonomous-driving
+- multimodal
+- motion-prediction
+- motion-generation
+- question-answering
+- benchmark
+- video
+language:
+- en
+size_categories:
+- 1M<n<10M
+---
+
+# MMHU: A Massive-Scale Multimodal Benchmark for Human Behavior Understanding
+
+This repository contains the MMHU (Massive-scale Multimodal benchmark for Human behavior Understanding) dataset, as presented in the paper [MMHU: A Massive-Scale Multimodal Benchmark for Human Behavior Understanding](https://huggingface.co/papers/2507.12463).
+
+MMHU is a large-scale benchmark designed for human behavior analysis, particularly relevant to autonomous driving systems. It features rich annotations crucial for understanding human actions in complex environments.
+
+**Key features of the dataset include:**
+
+*   **Scale:** Encompasses 57k human motion clips and 1.73M frames.
+*   **Annotations:** Provides detailed labels such as human motion and trajectories, text descriptions for human motions, human intention, and critical behavior labels relevant to driving safety.
+*   **Diversity:** Data is gathered from diverse sources, including established driving datasets (e.g., Waymo), in-the-wild videos from YouTube, and self-collected data.
+*   **Tasks:** Designed to benchmark various tasks, including:
+    *   Motion Prediction
+    *   Motion Generation
+    *   Human Behavior Question Answering
+
+**Project Page:** [https://MMHU-Benchmark.github.io](https://MMHU-Benchmark.github.io)