Add Nexar val/test annotations (749 videos, no videos redistributed)

README.md

@@ -1,3 +1,125 @@
 ---
-license: mit
+pretty_name: CAViAR
+license: other
+license_name: research-use
+license_link: https://github.com/nec-labs-ma/CAViAR/blob/main/LICENSE
+task_categories:
+  - visual-question-answering
+  - multiple-choice
+  - video-text-to-text
+language:
+  - en
+tags:
+  - accident
+  - dashcam
+  - vlm
+  - video
+  - responsibility
+  - nexar
+size_categories:
+  - n<1K
+configs:
+  - config_name: default
+    data_files:
+      - split: test
+        path: data/test.parquet
+dataset_info:
+  features:
+    - name: video_path
+      dtype: string
+    - name: qa_pairs
+      list:
+        - name: answer
+          dtype: string
+        - name: benchmark
+          dtype: string
+        - name: choices
+          sequence: string
+        - name: correct_answer
+          dtype: string
+        - name: correct_index
+          dtype: int64
+        - name: question
+          dtype: string
+  splits:
+    - name: test
+      num_examples: 749
 ---
+
+# CAViAR (val / test)
+
+**Causal Accident Video and Incident Analysis Repository** — Nexar validation/test annotations.
+
+This Hub dataset mirrors the public val/test split from [github.com/nec-labs-ma/CAViAR](https://github.com/nec-labs-ma/CAViAR): **749 videos**, **7,407 QA pairs**. **Videos are not included.** Obtain Nexar clips under the original Nexar license; `video_path` is a filename only (e.g. `nexar_00284.mp4`).
+
+Paper: *CAViAR: A Causal Video Dataset for Fine-Grained Accident Reasoning in Real-World Scenarios* ([arXiv:2608.19380](https://arxiv.org/abs/2608.19380)).
+
+## Load
+
+```python
+from datasets import load_dataset
+
+ds = load_dataset("sparshgarg57/CAViAR", split="test")
+print(ds[0]["video_path"], len(ds[0]["qa_pairs"]))
+```
+
+Canonical JSON (same as the GitHub file): [`data/test.json`](data/test.json).
+
+## Split
+
+| Split | Source | Role | Released here |
+|-------|--------|------|----------------|
+| Train | [CCD](https://github.com/Cogito2012/CarCrashDataset) | Fine-tuning | No |
+| Val / test | [Nexar](https://www.nexar.com/) | Held-out evaluation | Yes (annotations only) |
+
+## Tasks
+
+| `benchmark` | Type |
+|-------------|------|
+| Dense Captioning | Open |
+| Weather & Light | MCQ |
+| Road Conditions | MCQ |
+| Accident Type | MCQ |
+| Faulter Identification | Open (apparent at-fault agent) |
+| Victim Identification | Open (affected agent) |
+| Violation Identification | Open |
+| Accident Reason | Open |
+
+MCQ items include `choices`, `correct_answer`, and `correct_index`. Open-ended items leave those empty (`correct_index` = `-1`).
+
+## Videos
+
+Place Nexar mp4 files so names match `video_path`:
+
+```text
+nexar_00001.mp4
+...
+nexar_00284.mp4
+```
+
+Code, eval scripts, and ontology: [CAViAR GitHub](https://github.com/nec-labs-ma/CAViAR).
+
+Example clip (`nexar_00284`, T-bone):
+
+![Preview of nexar_00284.mp4](https://github.com/nec-labs-ma/CAViAR/raw/main/examples/nexar_00284.gif)
+
+Source video from Nexar (Moura et al., 2025). Short collision-window preview only; the Nexar corpus is not redistributed.
+
+## License and ethics
+
+Annotations are for academic research on accident understanding and rule-relevant multimodal reasoning. Raw Nexar (and CCD) videos remain under their original licenses.
+
+Labels are *apparent* responsibility cues from video evidence. They are **not** legal determinations of liability and must not be used for adjudication, insurance, enforcement, or decisions about identifiable individuals.
+
+## Citation
+
+```bibtex
+@article{garg2026caviar,
+  title={CAViAR: A Causal Video Dataset for Fine-Grained Accident Reasoning in Real-World Scenarios},
+  author={Garg, Sparsh and Chen, Yi-Wen and Aich, Abhishek and others},
+  journal={arXiv preprint arXiv:2608.19380},
+  year={2026}
+}
+```
+
+Please also cite **Nexar** when using the val/test videos (Moura et al., 2025).

data/test.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:b0b440473b7a99df0d72bec3c4cafceedc5dc1ae3b7476c98730838e28dbacea
+size 145727