Add dataset description, task category and paper link

README.md

@@ -2,7 +2,8 @@
 language:
 - en
 license: cc-by-sa-4.0
-
+task_categories:
+- video-text-to-text
 configs:
 - config_name: benchmark
   data_files:
@@ -18,4 +19,31 @@ configs:
     path: data/revplan.parquet
   - split: shape
     path: data/shape.parquet
----
+---
+
+# SAW-Bench: Situated Awareness in the Real World
+
+[Project Page](https://sawbench.github.io) | [Paper](https://huggingface.co/papers/2602.16682)
+
+SAW-Bench (Situated Awareness in the Real World) is a benchmark designed to evaluate the egocentric situated awareness of multimodal foundation models (MFMs). Unlike environment-centric benchmarks, SAW-Bench focuses on observer-centric relationships, requiring models to reason relative to an agent's viewpoint, pose, and motion.
+
+## Dataset Summary
+The dataset comprises 786 self-recorded videos captured using Ray-Ban Meta (Gen 2) smart glasses across diverse indoor and outdoor environments. It includes over 2,071 human-annotated question-answer pairs spanning six awareness tasks:
+
+- **Affordance**: Reasoning about possible actions in context.
+- **Direction**: Observer-centric spatial orientation.
+- **Localization**: Determining position relative to surroundings.
+- **Memory**: Reasoning over temporal events in the video.
+- **Revplan** (Reverse Planning): Inferring previous actions or intentions.
+- **Shape**: Understanding the geometric structure of the environment from the observer's perspective.
+
+## Citation
+If you use this dataset in your research, please cite the following paper:
+```bibtex
+@article{li2026learning,
+  title={Learning Situated Awareness in the Real World},
+  author={Li, Chuhan and Han, Ruilin and Hsu, Joy and Liang, Yongyuan and Dhawan, Rajiv and Wu, Jiajun and Yang, Ming-Hsuan and Wang, Xin Eric},
+  journal={arXiv preprint arXiv:2602.16682},
+  year={2026}
+}
+```