NonMatchingSplitsSizesError when loading partial dataset files

[QingGo]

Describe the bug

When loading only a subset of dataset files while the dataset's README.md contains split metadata, the system throws a NonMatchingSplitsSizesError . This prevents users from loading partial datasets for quick validation in cases of poor network conditions or very large datasets.

Steps to reproduce the bug

1. Use the Hugging Face datasets library to load a dataset with only specific files specified
2. Ensure the dataset repository has split metadata defined in README.md
3. Observe the error when attempting to load a subset of files

# Example code that triggers the error
from datasets import load_dataset

book_corpus_ds = load_dataset(
    "SaylorTwift/the_pile_books3_minus_gutenberg",
    name="default",
    data_files="data/train-00000-of-00213-312fd8d7a3c58a63.parquet",
    split="train",
    cache_dir="./data"
)

Error Message

Traceback (most recent call last):
  File "/Users/QingGo/code/llm_learn/src/data/clean_cc_bc.py", line 13, in <module>
    book_corpus_ds = load_dataset(
        "SaylorTwift/the_pile_books3_minus_gutenberg",
    ...
  File "/Users/QingGo/code/llm_learn/.venv/lib/python3.13/site-packages/datasets/utils/info_utils.py", line 77, in verify_splits
    raise NonMatchingSplitsSizesError(str(bad_splits))
datasets.exceptions.NonMatchingSplitsSizesError: [{'expected': SplitInfo(name='train', num_bytes=106199627990.47722, num_examples=192661, shard_lengths=None, dataset_name=None), 'recorded': SplitInfo(name='train', num_bytes=454897326, num_examples=905, shard_lengths=None, dataset_name='the_pile_books3_minus_gutenberg')}]

Expected behavior

When loading partial dataset files, the system should:

1. Skip the NonMatchingSplitsSizesError validation, OR
2. Only log a warning message instead of raising an error

Environment info

• datasets version: 4.3.0
• Platform: macOS-15.7.1-arm64-arm-64bit-Mach-O
• Python version: 3.13.2
• huggingface_hub version: 0.36.0
• PyArrow version: 22.0.0
• Pandas version: 2.3.3
• fsspec version: 2025.9.0

[QingGo]

While using verification_mode='no_checks' parameter in load_dataset() can bypass this validation, this solution is not intuitive or convenient for most users, especially those who are not familiar with all the parameters of the load_dataset() function.

book_corpus_ds = load_dataset(
    "SaylorTwift/the_pile_books3_minus_gutenberg",
    name="default",
    data_files="data/train-00000-of-00213-312fd8d7a3c58a63.parquet",
    split="train",
    cache_dir="./data",
    verification_mode='no_checks'
)

[ArjunJagdale]

Thanks for the report and reproduction steps @QingGo
@lhoestq which one of the following looks like a nicer way to handle this?

1] Skip split-size validation entirely for partial loads
If the user passes data_files manually and it represents only a subset, then verify_splits() should simply not run, or skip validation only for that split.

2] Replace the error with a warning

3] Automatically detect partial-load cases(i mean we can try this out!)

Assume this,
If data_files is provided AND
the number of provided files ≠ number of expected files in metadata,
then treat it as a partial load and disable strict verification.