| # SCP-Bench |
|
|
| SCP-Bench is a benchmark for Spatial Causal Prediction. |
|
|
| ## v2 release |
|
|
| During our pre-release review, we identified issues in some samples involving annotations, temporal boundaries, or supporting video evidence. This revised release corrects affected annotations and video boundaries and removes selected samples, retaining 2,427 samples with their original IDs. The original v1 release remains available for reproducibility. |
|
|
| ## Versions |
|
|
| - [v1: 2,500 samples](https://huggingface.co/datasets/GuangsTrip/SCP-Bench/tree/v1) — original release. |
| - [v2: 2,427 samples](https://huggingface.co/datasets/GuangsTrip/SCP-Bench/tree/v2) — revised release. |
|
|
| Download `test.jsonl` and `SCPBench.zip` from the same version. |
|
|
| ## Usage |
|
|
| Extract `SCPBench.zip` alongside `test.jsonl`. |
| Media paths in the annotations are relative to this directory. |
|
|
| Use `input_video` and, when provided, `ref_image` as visual inputs. |
| `full_video` is provided for reference, not as prediction input. |
| `ground_truth` contains the correct option text. |
|
|
