[BarsamErshadi]

What does this PR do?

When upload_dataset uploads a directory image by image, it infers each image's split in _decide_split, which looks for valid, train and test anywhere in the lowercased relative path, including class folders and the filename. So a class folder or filename can silently override the real split directory: with a classification class called train, images under test/train/ are uploaded to train. The split= argument forces one split for the whole upload, so it can't repair a mixed dataset. Zip uploads (a .zip path or use_zip_upload=True) return before this code runs, so this PR doesn't affect them.

With this change, _decide_split works in two steps:

1. It takes the topmost directory whose whole name is a split, ignoring case: train, training, valid, validation, test or testing, optionally followed by set, data or images (valid_set, train-data). So test/train/ is test and validation/train/ is valid.
2. If there is no such directory, it looks for a split name at the start of a word anywhere in the path, in the same valid, train, test order as before. A change of case also starts a word, so valid_001.jpg, my_test_set/, myTest/ and VOCtest/ keep working, while latest/ and invalid/ stop matching. The default is still train.

A folder that only contains a split word, like test-project-1/ or test_dataset/, doesn't count in step 1. That keeps test-project-1/train/ and test-project-1/valid/ apart instead of sending both to test.

Minimal repro:

import os, tempfile
from PIL import Image
from roboflow.util.folderparser import parsefolder

with tempfile.TemporaryDirectory() as d:
    for path in ["test/train/1.jpg", "test/bus/1.jpg", "train/invalid_ticket.jpg", "unsplit/latest.jpg"]:
        p = os.path.join(d, path)
        os.makedirs(os.path.dirname(p), exist_ok=True)
        Image.new("RGB", (4, 4)).save(p)
    for i in parsefolder(d, is_classification=True)["images"]:
        print(i["file"], "split=" + i["split"], "label=" + i["annotationfile"]["classification_label"])

Before:

/test/bus/1.jpg split=test label=bus
/test/train/1.jpg split=train label=train
/train/invalid_ticket.jpg split=valid label=train
/unsplit/latest.jpg split=test label=unsplit

After:

/test/bus/1.jpg split=test label=bus
/test/train/1.jpg split=test label=train
/train/invalid_ticket.jpg split=train label=train
/unsplit/latest.jpg split=train label=unsplit

I also called Workspace.upload_dataset with a stubbed project, on a folder with test/train/1.jpg, test/bus/1.jpg, validation/train/1.jpg and training/test/1.jpg. The split passed to upload_image for /test/train/1.jpg was train before and test after. The other three kept the splits they had before: test, valid and train.

Related Issue(s): None that I found. Open PRs #575, #490 and #421 also touch folderparser.py, but none of their hunks touch _decide_split.

Type of Change

• Bug fix (non-breaking change that fixes an issue)

The split of some images does change, which is the point of the fix. Additional Context has a table of what changed in a sample of paths, including the changes that could be regressions.

Testing

• I have tested this change locally
• I have added/updated tests for this change

Test details:

I added seven tests to tests/util/test_folderparser.py. They cover:

• a split directory winning over later path text
• an inner split name not overriding an outer directory like validation/
• wrapper folders like test-project-1/ not flattening the splits
• case-insensitive directory matching
• split names inside other words (latest, invalid, contest)
• a case change counting as a word start (myTest, VOCtest_06-Nov-2007, but not LATEST)
• the fallback for flat names and folders such as validation/

Six of them fail on the unfixed code. The fallback test passes both before and after, because it guards behaviour that has to stay the same.

• Fixtures: I dumped {file: split} for all 8 datasets in tests/datasets (72 images) before and after. The two dumps are byte-identical.
• tests/util: 158 tests, OK.
• Full suite (unittest discovery over tests/, with non-loopback connections blocked): 1285 tests, 1 skipped, 1 error. The error is tests.cli.test_train_handler.TestTrainStart.test_start_with_all_options, which tries a real outbound HTTPS connection. It errors the same way with this fix removed, so it is caused by my network block rather than this change. I did not run it with network access.
• Hooks: pre-commit passed on the changed files, including bandit, ruff format and ruff. mypy on folderparser.py also passed.
• Limits: pytest isn't installed in my environment, so I used unittest, which is what CONTRIBUTING.md describes. I ran everything on macOS only and did not test against the real Roboflow API.

Checklist

• My code follows the style guidelines of this project
• I have performed a self-review of my own code
• I have commented my code where necessary, particularly in hard-to-understand areas
• My changes generate no new warnings or errors
• I have updated the documentation accordingly (if applicable) (no docs describe this inference, so nothing needed changing)

Additional Context

These are the behaviour changes for you to judge. I computed the old and new splits for 76 sample paths, and 32 changed. All 32 are in this table. It's a sample, so other paths can change too.

Path	Before	After	Note
test/train/a.jpg, test/valid/a.jpg, dataset/test/train/a.jpg	train, valid, train	test	the bug
train/valid/a.jpg, train/valid_001.jpg, train/invalid_ticket.jpg	valid	train	the bug
test/train_001.jpg	train	test	the bug
test_set/train/a.jpg, test_set/valid/a.jpg	train, valid	test	the bug
testing/train/a.jpg, testing/train_001.jpg	train	test	the bug
train_test_split/test/a.jpg	train	test	the bug
invalid/a.jpg, INVALID/a.jpg, Invalid/a.jpg	valid	train	the bug
latest/a.jpg, LATEST/a.jpg, Latest/a.jpg, contest/a.jpg	test	train	the bug
backtest/a.jpg	test	train	word-boundary trade-off
mytest/a.jpg, xtest/a.jpg, MYTEST/a.jpg	test	train	trade-off: possible regression
myvalid/a.jpg	valid	train	trade-off: possible regression
trainvalid/a.jpg	valid	train	trade-off: possible regression
testtrain/a.jpg	train	test	trade-off: possible regression
testing/valid/a.jpg	valid	test	trade-off: outer directory wins
training/valid/a.jpg	valid	train	trade-off: outer directory wins
test/fixtures/train/a.jpg	train	test	trade-off: topmost wins
training_stuff/test/a.jpg, train_v2/test/a.jpg, train2017/test/a.jpg	train	test	trade-off: possible regression

• Run-together names: a split name that follows a letter of the same case is no longer recognised, so mytest/, xtest/ and MYTEST/ now land in train. That rule is what stops latest and invalid from matching. Separated and mixed-case forms still work: my_test, my-test, myTest, imagesValid and VOCtest_06-Nov-2007. A dataset that relies on all-lowercase run-together names would change silently.
• Case changes: because a case change starts a word, laTest/ and inValid/ still match, as they did before.
• Unrecognised outer folders: the list of split directory names is short on purpose, and an outer folder that isn't on it loses to an exact name inside it. So train_v2/test/ and train2017/test/ go to test, which is wrong if test is a class. The same rule is what keeps test-project-1/train/ in train, and a per-image rule can't tell those two apart.
• Wrapper directories: if a folder named like a split sits above the real split directory, the outer one now wins, so testing/valid/ is test and training/valid/ is train. The path is relative to the uploaded folder, so this only matters when that wrapper is inside the upload.
• Still unchanged:
  • validation/train/a.jpg and validation/test/a.jpg stay valid, and training/test/a.jpg stays train.
  • my_test_set/train/a.jpg and test-project-1/train/a.jpg stay train.
  • val/images/b.jpg still goes to train, and so does val2017/train/a.jpg. I left treating val as a split name for you to decide.