update dataset card

README.md

@@ -111,7 +111,9 @@ configs:
 
 # BDD100K Train
 
-This dataset is derived from the [BDD100K](https://www.bdd100k.com/) driving video dataset. It contains 69,800 training sequences of 15 frames each.
+This dataset was created for [Dedelayed: Deleting Remote Inference Delay via On-Device Correction](https://arxiv.org/pdf/2510.13714) (CVPR 2026). Code is available at [InterDigitalInc/dedelayed](https://github.com/InterDigitalInc/dedelayed).
+
+The underlying data is derived from the [BDD100K](https://www.bdd100k.com/) driving video dataset. It contains 69,800 training sequences of 15 frames each.
 
 ## Usage
 
@@ -128,12 +130,6 @@ dataset = datasets.DatasetDict({
 
 For an example training `collate_fn`, see the [reference training notebook](https://github.com/InterDigitalInc/dedelayed/blob/papers/2026-cvpr-dedelayed/bdd100k_mixed_res.ipynb).
 
-## Source
-
-This dataset was created for [Dedelayed: Deleting Remote Inference Delay via On-Device Correction](https://arxiv.org/pdf/2510.13714) (CVPR 2026). Code is available at [InterDigitalInc/dedelayed](https://github.com/InterDigitalInc/dedelayed).
-
-The underlying data is derived from BDD100K. Please see the [license](#license) and [citation](#citation) below.
-
 ## License
 
 This dataset is distributed under the BDD100K license: