Add 150 benchmark tasks with initial observations

README.md

@@ -1,3 +1,99 @@
 ---
 license: apache-2.0
+pretty_name: GTAVBench
+language:
+- en
+tags:
+- gui-agents
+- game-environments
+- multimodal
+configs:
+- config_name: default
+  data_files:
+  - split: test
+    path: data/test-*.parquet
 ---
+
+# GTAVBench
+
+This dataset contains the frozen 150-task GTAVBench benchmark across 11 task
+categories: 103 native 3D tasks and 47 image-based shooting/sniping tasks.
+The executable environments, action interfaces, evaluators, task definitions,
+and credits are in [the GTASuite repository](https://github.com/ljbro/GTASuite).
+
+## Fields
+
+| Field | Type | Description |
+| --- | --- | --- |
+| `id` | string | Public task ID, `cXX-nYY`. |
+| `category` | string | Human-readable task category. |
+| `instruction` | string | Complete English instruction used in evaluation. |
+| `initial_observation` | Image | The 1280 x 720 screenshot before the first agent action. |
+| `environment_type` | string | `3d` or `image`. |
+| `movement` | list of strings | Movement modes involved in the task; see below. |
+
+## Movement vocabulary
+
+The vocabulary follows the paper's Movement parameter:
+`walk`, `drive`, `swim`, `dive`, `boat`, `plane`, `helicopter`.
+Multiple modes are represented as a list, for example `["walk", "drive"]`
+for a task that starts on foot and requires finding and driving a vehicle.
+The list describes the benchmark task's movement modes; it is not a replay
+sequence or a generator sampling configuration.
+
+Cars, motorcycles, vans, trucks, and ATVs use `drive`; their concrete vehicle
+requirements are preserved in the instruction. Vertical traversal uses `walk`,
+including stairs, ladders, and climbing. Watercraft use `boat`; the submarine
+task additionally uses `dive` for underwater navigation. Boarding a boat is
+included in `boat`, even when the task ends immediately after boarding.
+Tasks starting on foot before boarding an aircraft include `walk` and `plane`
+or `helicopter`. Image-based shooting and sniping use `[]`, since locomotion
+does not apply; no extra `stationary` vocabulary item is introduced.
+
+## Initial observations
+
+Screenshots are reused from evaluation observations at game elapsed time zero,
+before the first agent action. They are not later trajectory frames. Sniping
+observations show the initial unscoped view. Screenshot bytes are preserved
+without resizing or recompression and embedded directly in Parquet with the
+Hugging Face `Image` feature. Screenshots may contain the original in-game HUD
+and, for harness observations, the instruction overlay shown to the agent.
+
+## Task distribution
+
+| Category | Tasks |
+| --- | ---: |
+| Road Navigation | 20 |
+| On-foot Navigation | 16 |
+| Following | 6 |
+| Interception | 13 |
+| Evasion | 16 |
+| Aquatic Navigation | 11 |
+| Vertical Traversal | 5 |
+| Aerial Navigation | 8 |
+| Indoor Navigation | 8 |
+| Shooting | 24 |
+| Sniping | 23 |
+
+The single `test` split contains all 150 benchmark tasks. RL training tasks are
+not included in this release. Time limits, round budgets, model outputs,
+reasoning traces, evaluation scores, and private experiment metadata are not
+dataset columns. This task catalog and its screenshots accompany the runnable
+environment; a screenshot alone is not a complete executable game state.
+
+## Loading
+
+```python
+from datasets import load_dataset
+
+dataset = load_dataset("OpenMOSS-Team/GTAVBench", split="test")
+image = dataset[0]["initial_observation"]
+```
+
+## Acknowledgements
+
+The native 3D environments are built on
+[GTA_SZ](https://github.com/linranff/GTA_SZ). Task construction also draws on
+OffMapBench and GTA V gameplay. See the GTASuite repository's
+[acknowledgements](https://github.com/ljbro/GTASuite/blob/main/ACKNOWLEDGEMENTS.md)
+and asset-specific notices for credits, provenance, and licenses.

data/test-00000-of-00001.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b7af48103661a7506a91b5a73d208b21c83d5a3e18106b8caf58cbcb05711b6f
+size 79784903