From ff766b835fa19039043a4cb4ebc37ee03df86cc1 Mon Sep 17 00:00:00 2001
From: Deepseko <Bruvbtw@gmail.com>
Date: Wed, 4 Nov 2026 05:34:35 -0500
Subject: [PATCH 1/2] fix(folderparser): take the split from directory names
 before substrings

_decide_split looked for "valid", "train" and "test" anywhere in the
lowercased relative path, so a class folder or filename could override
the split directory. In a classification dataset, test/train/1.jpg was
uploaded to train and train/invalid_ticket.jpg to valid, and
unsplit/latest.jpg went to test because of "latest". Nothing warned about
it, and the split= argument can't repair a mixed dataset.

Take the topmost directory whose whole name is a split (test/,
validation/, train_set/), so validation/train/ stays valid while a wrapper
like test-project-1/train/ stays train. Otherwise fall back to a match at
the start of a word, in the old valid, train, test order, treating a
change of case as a word start, so valid_001.jpg, my_test_set/, myTest/
and VOCtest/ keep working and latest/ and invalid/ stop matching.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---
 roboflow/util/folderparser.py   |  31 +++++---
 tests/util/test_folderparser.py | 137 ++++++++++++++++++++++++++++++++
 2 files changed, 159 insertions(+), 9 deletions(-)

diff --git a/roboflow/util/folderparser.py b/roboflow/util/folderparser.py
index 047cdaee..c79f127c 100644
--- a/roboflow/util/folderparser.py
+++ b/roboflow/util/folderparser.py
@@ -10,6 +10,16 @@
 IMAGE_EXTENSIONS = {".jpg", ".jpeg", ".png", ".bmp", ".tiff", ".tif", ".avif", ".heic"}
 ANNOTATION_EXTENSIONS = {".txt", ".json", ".xml", ".csv", ".jsonl"}
 LABELMAPS_EXTENSIONS = {".labels", ".yaml", ".yml"}
+SPLIT_NAMES = ("valid", "train", "test")
+# A directory that is a split as a whole: test/, validation/, train_set/. Not my_test/ or test_dataset/.
+SPLIT_DIR = re.compile(
+    r"(?:(?P<valid>valid(?:ation)?)|(?P<train>train(?:ing)?)|(?P<test>test(?:ing)?))(?:[-_ ]?(?:set|data|images))?",
+    re.IGNORECASE,
+)
+# A split name at the start of a word, where a change of case also starts one: my_test, myTest, VOCtest, not latest.
+SPLIT_WORDS = {
+    name: re.compile(rf"(?:(?<![A-Za-z])|(?<=[a-z])(?=[A-Z])|(?<=[A-Z])(?=[a-z]))(?i:{name})") for name in SPLIT_NAMES
+}
 
 
 def _patch_sep(filename):
@@ -381,15 +391,18 @@ def _load_labelmaps(folder, labelmaps):
 
 def _decide_split(images):
     for i in images:
-        fullkey = i["fullkey"]
-        if "valid" in fullkey:
-            i["split"] = "valid"
-        elif "train" in fullkey:
-            i["split"] = "train"
-        elif "test" in fullkey:
-            i["split"] = "test"
-        else:
-            i["split"] = "train"
+        i["split"] = _split_for_path(i["dirname"], i["file"])
+
+
+def _split_for_path(dirname, file):
+    # A split directory wins, topmost first, so class folders and filenames can't override it.
+    for segment in dirname.split("/"):
+        match = SPLIT_DIR.fullmatch(segment)
+        if match:
+            return match.lastgroup
+    # Otherwise keep the old valid, train, test order, for names like valid_001.jpg and my_test_set/.
+    stem = os.path.splitext(file)[0]
+    return next((name for name in SPLIT_NAMES if SPLIT_WORDS[name].search(stem)), "train")
 
 
 def _infer_classification_labels_from_folders(images):
diff --git a/tests/util/test_folderparser.py b/tests/util/test_folderparser.py
index 4f9ddb5b..68d0e2f5 100644
--- a/tests/util/test_folderparser.py
+++ b/tests/util/test_folderparser.py
@@ -205,6 +205,143 @@ def test_coco_root_annotation_matches_images_in_subdirs(self):
             self.assertEqual(len(ann_data["annotations"]), 1, "Should have one annotation")
             self.assertEqual(ann_data["annotations"][0]["bbox"], [10, 20, 100, 200])
 
+    def test_split_directory_wins_over_later_path_text(self):
+        # A class folder or filename that contains a split name must not override the split directory.
+        self.assertEqual(
+            _splits_for(
+                [
+                    "test/train/1.jpg",
+                    "test/bus/1.jpg",
+                    "test/valid/1.jpg",
+                    "train/invalid_ticket.jpg",
+                    "train/valid/1.jpg",
+                    "test/images/train_001.jpg",
+                ]
+            ),
+            {
+                "/test/train/1.jpg": "test",
+                "/test/bus/1.jpg": "test",
+                "/test/valid/1.jpg": "test",
+                "/train/invalid_ticket.jpg": "train",
+                "/train/valid/1.jpg": "train",
+                "/test/images/train_001.jpg": "test",
+            },
+        )
+
+    def test_inner_split_name_does_not_override_outer_split_directory(self):
+        # The motivating case again, with a split folder that isn't literally train, valid or test.
+        self.assertEqual(
+            _splits_for(
+                [
+                    "validation/train/1.jpg",
+                    "VALIDATION_SET/Train/1.jpg",
+                    "validation/test/1.jpg",
+                    "valid_set/train/1.jpg",
+                    "training/test/1.jpg",
+                    "training_set/valid/1.jpg",
+                    "train_data/test/1.jpg",
+                    "test_set/train/1.jpg",
+                ]
+            ),
+            {
+                "/validation/train/1.jpg": "valid",
+                "/VALIDATION_SET/Train/1.jpg": "valid",
+                "/validation/test/1.jpg": "valid",
+                "/valid_set/train/1.jpg": "valid",
+                "/training/test/1.jpg": "train",
+                "/training_set/valid/1.jpg": "train",
+                "/train_data/test/1.jpg": "train",
+                "/test_set/train/1.jpg": "test",
+            },
+        )
+
+    def test_wrapper_folder_with_a_split_word_does_not_flatten_the_splits(self):
+        # A project folder such as test-project-1/ holds the real splits, so it must not take them over.
+        self.assertEqual(
+            _splits_for(
+                [
+                    "test-project-1/train/1.jpg",
+                    "test-project-1/valid/1.jpg",
+                    "test_dataset/train/1.jpg",
+                    "my_test_set/train/1.jpg",
+                    "train_test_split/test/1.jpg",
+                    "train_test_split/valid/1.jpg",
+                ]
+            ),
+            {
+                "/test-project-1/train/1.jpg": "train",
+                "/test-project-1/valid/1.jpg": "valid",
+                "/test_dataset/train/1.jpg": "train",
+                "/my_test_set/train/1.jpg": "train",
+                "/train_test_split/test/1.jpg": "test",
+                "/train_test_split/valid/1.jpg": "valid",
+            },
+        )
+
+    def test_split_recognises_a_case_change_as_a_word_start(self):
+        self.assertEqual(
+            _splits_for(
+                [
+                    "myTest/1.jpg",
+                    "imagesValid/1.jpg",
+                    "VOCtest_06-Nov-2007/JPEGImages/1.jpg",
+                    "LATEST/1.jpg",
+                    "Contest/1.jpg",
+                    "INVALID/1.jpg",
+                ]
+            ),
+            {
+                "/myTest/1.jpg": "test",
+                "/imagesValid/1.jpg": "valid",
+                "/VOCtest_06-Nov-2007/JPEGImages/1.jpg": "test",
+                "/LATEST/1.jpg": "train",
+                "/Contest/1.jpg": "train",
+                "/INVALID/1.jpg": "train",
+            },
+        )
+
+    def test_split_directory_match_ignores_case(self):
+        self.assertEqual(_splits_for(["TEST/train/1.jpg"]), {"/TEST/train/1.jpg": "test"})
+
+    def test_split_ignores_split_names_inside_other_words(self):
+        self.assertEqual(
+            _splits_for(["unsplit/latest.jpg", "invalid/1.jpg", "contest/1.jpg"]),
+            {"/unsplit/latest.jpg": "train", "/invalid/1.jpg": "train", "/contest/1.jpg": "train"},
+        )
+
+    def test_split_falls_back_to_names_without_a_split_directory(self):
+        self.assertEqual(
+            _splits_for(
+                [
+                    "train_001.jpg",
+                    "valid_001.jpg",
+                    "test_001.jpg",
+                    "validation/a.jpg",
+                    "training/a.jpg",
+                    "my_test_set/a.jpg",
+                    "other/a.jpg",
+                ]
+            ),
+            {
+                "/train_001.jpg": "train",
+                "/valid_001.jpg": "valid",
+                "/test_001.jpg": "test",
+                "/validation/a.jpg": "valid",
+                "/training/a.jpg": "train",
+                "/my_test_set/a.jpg": "test",
+                "/other/a.jpg": "train",
+            },
+        )
+
+
+def _splits_for(relpaths):
+    with tempfile.TemporaryDirectory() as tmpdir:
+        for relpath in relpaths:
+            path = os.path.join(tmpdir, relpath)
+            os.makedirs(os.path.dirname(path), exist_ok=True)
+            open(path, "wb").close()
+        return {i["file"]: i["split"] for i in folderparser.parsefolder(tmpdir)["images"]}
+
 
 def _assertJsonMatchesFile(actual, filename):
     with open(filename) as file:

From 42b97319df3b8fb9b79bec08150fcbc75ceb2e56 Mon Sep 17 00:00:00 2001
From: Deepseko <Bruvbtw@gmail.com>
Date: Wed, 4 Nov 2026 06:30:59 -0500
Subject: [PATCH 2/2] docs(changelog): add entry for this fix

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
---
 CHANGELOG.md | 12 ++++++++++++
 1 file changed, 12 insertions(+)

diff --git a/CHANGELOG.md b/CHANGELOG.md
index 3de8e4fd..dd0275c2 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -34,6 +34,18 @@ All notable changes to this project will be documented in this file.
   YOLOv10. Existing models remain available for inference, and YOLO dataset
   formats remain supported ([#572](https://github.com/roboflow/roboflow-python/pull/572)).
 
+### Fixed
+
+- Directory uploads with `upload_dataset` now take each image's split from the
+  topmost directory whose whole name is a split (`test/`, `validation/`,
+  `train_set/`), so a class folder or filename inside it can't override it:
+  `test/train/1.jpg` goes to test, not train. Without such a directory, a split
+  name only counts at the start of a word, so `latest/` and `invalid/` no longer
+  match, while `my_test/`, `myTest/` and `valid_001.jpg` still do. Run-together
+  names like `mytest/` now go to train, and an unrecognised outer folder loses to
+  an exact name inside it, so `train_v2/test/` goes to test
+  ([#576](https://github.com/roboflow/roboflow-python/pull/576)).
+
 ## 1.7.3
 
 ### Added