`load_dataset` `revision` param not respected when fetching from cache

[Scott-Simmons]

Describe the bug

datasets.load_dataset revision semantics are a bit inconsistent when the dataset is not found on the huggingface hub. When fetching the latest cached version of the dataset, the revision argument is ignored, so long as any cached versions of the dataset already exist in the HF cache.

Steps to reproduce the bug

import datasets
datasets.load_dataset(
    "sentientfutures/ahb",
    "dimensions",
    split="train",
    revision="main"
)
# would expect some error to raise here
datasets.load_dataset(
    "sentientfutures/ahb",
    "dimensions",
    split="train",
    revision="invalid_revision"
)

Expected behavior

On the second call to datasets.load_dataset in the 'steps to reproduce the bug' example, expect something like:

raise DatasetNotFoundError(
datasets.exceptions.DatasetNotFoundError: Revision 'invalid_revision' doesn't exist for dataset 'sentientfutures/ahb' on the Hub.

Environment info

• datasets version: 4.4.1
• Platform: Linux-6.2.0-39-generic-x86_64-with-glibc2.37
• Python version: 3.12.12
• huggingface_hub version: 0.36.0
• PyArrow version: 22.0.0
• Pandas version: 2.2.3
• fsspec version: 2025.9.0

[Scott-Simmons]

This might be better placed as a feature request not a bug, since the logging Using the latest cached version of the dataset since sentientfutures/ahb couldn't be found on the Hugging Face Hub is clear.

[Scott-Simmons]

#7929