Add dataset card, link to paper and code
Browse filesThis PR adds a proper dataset card for StateBench, including the task category `text-to-video`, a link to the paper (https://huggingface.co/papers/2609.03673), and a link to the code repository.
README.md
CHANGED
|
@@ -1,3 +1,21 @@
|
|
| 1 |
---
|
| 2 |
license: apache-2.0
|
|
|
|
|
|
|
| 3 |
---
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
---
|
| 2 |
license: apache-2.0
|
| 3 |
+
task_categories:
|
| 4 |
+
- text-to-video
|
| 5 |
---
|
| 6 |
+
|
| 7 |
+
# StateBench
|
| 8 |
+
|
| 9 |
+
StateBench is a benchmark for world-state reasoning in video continuation, introduced in the paper [Do Video Generators Track the World Across Segments? A Benchmark and Method for World-State Reasoning in Video Continuation](https://huggingface.co/papers/2609.03673).
|
| 10 |
+
|
| 11 |
+
Code: [https://github.com/AMAP-ML/StateAgent](https://github.com/AMAP-ML/StateAgent)
|
| 12 |
+
|
| 13 |
+
The benchmark contains **200 cross-segment continuation tasks** across three difficulty levels:
|
| 14 |
+
|
| 15 |
+
| Level | Count | Description |
|
| 16 |
+
|-------|-------|-------------|
|
| 17 |
+
| `past_visible` | 85 | The target object was visible in a prior frame — tests temporal memory. |
|
| 18 |
+
| `occluded_process` | 65 | The state change occurs while the object is occluded — tests process reasoning. |
|
| 19 |
+
| `complex_transition` | 50 | Multiple state transitions or conflicting cues — tests compositional reasoning. |
|
| 20 |
+
|
| 21 |
+
Each task includes metadata such as `id`, `difficulty`, `target_object`, `shots`, `expected_state`, `checklist`, and `reference_frames`.
|