`set_format` resets features

[plutonium-239]

Describe the bug

Calling dataset.set_format('...') resets the informative features such as Array2D etc. back into List(List(Value)) and so on.

Steps to reproduce the bug

>>> dataset.features
{'lld': Array2D(shape=(None, 26), dtype='float32')}

>>> dataset.set_format('torch')
>>> dataset.features
{'lld': List(List(Value('float32')))}

Expected behavior

The feature information is not lost during set_format, and the feature remains Array2D with the correct shape and dtype in the above example.

Environment info

• datasets version: 4.8.2
• Platform: Linux-6.8.0-78-generic-x86_64-with-glibc2.39
• Python version: 3.13.11
• huggingface_hub version: 0.36.2
• PyArrow version: 23.0.1
• Pandas version: 2.3.3
• fsspec version: 2026.2.0

[ta0a2000t]

couldn't replicate with:

Platform: macos
python: 3.10.9
datasets: 4.8.4
huggingface_hub: 0.36.2
pyarrow: 23.0.1
pandas: 2.3.3
fsspec: 2026.2.0

import datasets
from datasets import Array2D, Dataset
import numpy as np

# Create a dataset with Array2D feature
data = {"lld": [np.zeros((5, 26), dtype="float32").tolist() for _ in range(10)]}

features = datasets.Features({"lld": Array2D(shape=(None, 26), dtype="float32")})
dataset = Dataset.from_dict(data, features=features)

print("Before set_format:")
print(dataset.features)

dataset.set_format("torch")

print("\nAfter set_format('torch'):")
print(dataset.features)