Dataset images with incorrect orientation

[Phonicavi]

Hi,

I've noticed that several images in huggingface datasets are rotated (90°/180°/270°), is that a bug or a feature?

Dataset used:
https://huggingface.co/datasets/Cusyoung/SpaCE-10

Rotations observed on a random sample.
Code used:

from PIL import Image, ImageOps
from datasets import load_dataset
import io

def decode_bytes(x: str):
    return Image.open(io.BytesIO(x))

def visualize(row: Dict) -> None:
    images = [decode_bytes(x) for x in row['image']]
    row.pop('image')
    print(json.dumps(row, indent=2))
    plt.close()
    _, ax = plt.subplots(4, 2, figsize=(16, 20))
    for idx, img in enumerate(images):
        ax[idx % 4][idx // 4].imshow(img)
    plt.show()

ds = load_dataset("/path/to/huggingface/datasets/Cusyoung/SpaCE-10/single-choice/sp", split='test')

sample = ds[76]
visualize(sample)

Also, img.getexif() shows empty, and ImageOps.exif_transpose(img) does not work for missing exif image.
How can I get the original EXIF to verify orientation?

[Phonicavi]

e.g.

{
  "index": 76,
  "question": "Which of the following descriptions are incorrect for keeping the room clean and tidy?",
  "answer": "B",
  "A": "Starting from the entrance, move right front to corner of the wall, then take down the painting on the wall, clean it, and hang it back up.",
  "B": "Starting from the entrance, move left front to tv, then wipe the tv clean with a cloth because it is not playing any programs at the moment.",
  "C": "Starting from the entrance, move front to small table, then vacuum the brown carpet under the table clean.",
  "D": "Starting from the entrance, move right front to black sofa, then arrange the pink pillow on the sofa properly.",
  "category": "sp_single"
}