`Dataset.split` silently accepts split ratios outside [0, 1]

[archie0732]

Search before asking

• I have searched the issues and discussions and found no similar bug report.

Bug

[Bug]: Dataset.split silently accepts split ratios outside [0, 1]

Bug

DetectionDataset.split() and ClassificationDataset.split() silently accept
finite ratios below 0 or above 1. They delegate to train_test_split, which
calculates int(len(data) * train_ratio) and uses Python slicing without
validating the ratio.

For a 10-image dataset, split_ratio=-0.2 produces an 8/2 split, while
split_ratio=1.2 (or an accidental percentage such as 80) produces a 10/0
split. A configuration error therefore looks like a successful split and can
leave no held-out data or produce an unintended partition.

This is an input-validation issue; valid ratios do not show a partitioning bug.

Expected behavior

Raise a clear ValueError for ratios outside the inclusive interval [0, 1]
and for non-finite values. Preserve the existing valid behavior of 0.0 and
1.0, which is explicitly covered by current tests, including empty datasets.

NaN and infinity already raise conversion errors; those are secondary error-
message consistency cases, not examples of silently accepted invalid inputs.

Minimal Reproducible Example

import supervision as sv

paths = [f"image_{i}.jpg" for i in range(10)]
dataset = sv.DetectionDataset(
    classes=["object"],
    images=paths,
    annotations={path: sv.Detections.empty() for path in paths},
)

# No actual image files, model, or GPU are needed: split uses image paths.
for ratio in [-0.2, 1.2, 80.0]:
    train, test = dataset.split(split_ratio=ratio, shuffle=False)
    print(ratio, len(train), len(test))

Output:

-0.2 8 2
1.2 10 0
80.0 10 0

Environment and verification

• Windows, Python 3.12.14, NumPy 2.5.3.
• Reproduced with the PyPI release supervision==0.30.4.
• Also reproduced against develop commit
  d79110b149c26d33b4aed6d2d14436dfaa714e5c (source imported via PYTHONPATH).
• Ran 234 cases per version: three APIs, three dataset sizes (0, 1, 10), two
  shuffle settings, and 13 ratios. Valid ratios preserved partitions and input
  ordering; calls did not change the global random state.
• Existing upstream split tests: 18 passed.
• Six proposed out-of-range validation regression tests: all fail with
  DID NOT RAISE ValueError on unmodified develop.
• OpenCV is not installed; Supervision selected its NumPy fallback. These
  tests do not decode images or call image-processing functions.

Proposed scope

• Validate the ratio in the shared helper before shuffling or slicing.
• Document the accepted inclusive range and exception.
• Add regression tests for both public dataset classes and the helper, covering
  negative/greater-than-one/non-finite ratios, empty input, both shuffle modes,
  and unchanged behavior for valid ratios including 0 and 1.
• Avoid adding new dataset-splitting APIs or changing rounding behavior.

Related work checked

• PR fix(dataset): stop split mutation, determinize VOC, guard collisions #2394 fixed split mutation and global random-state effects.
• PR test: cover public API gaps and dataset split #2399 added split tests, including support for ratios 0 and 1.
• PR fix: remaining review findings in dataset, docs, and tests #2416 addressed in-memory dataset handling and other dataset issues.
• PR Add Efficient COCO Dataset Splitting Utilities #1946 proposed separate COCO splitting utilities and was closed as a
  duplicate of existing dataset functionality. Its percentage validation is in
  a new, unmerged API, not a fix to the current Dataset.split() helper.

I did not find an existing issue or PR specifically addressing this validation
gap in the searches performed on 2026-09-21.

Would a focused validation fix preserving the existing [0, 1] behavior be
welcome?

Environment

• Supervision:
• Python:
• OS:

Minimal Reproducible Example

No response

Are you willing to submit a PR?

• Yes I'd like to help by submitting a PR!