fix(yolo): correct data.yaml paths for all YOLO formats (#240)

[davidnichols-ops]

Summary

Fixes #240 — Incorrect Data Path in YOLOv8 Dataset Configuration (5 comments, confirmed regression from v1.1.40 → v1.2.16, reported by multiple users).

When downloading a YOLO dataset via version.download("yolov8"), the generated data.yaml contained incorrect relative paths:

train: ../train/images
val: ../valid/images
test: ../test/images

The ../ prefix is wrong — the train/, valid/, test/ directories are siblings of data.yaml, not in the parent directory. This broke YOLO training out of the box.

Root cause — four bugs in __reformat_yaml

1. yolov8 path fix was gated on ultralytics==8.0.196 — for any other version (or when ultralytics wasn't installed at all) paths remained broken. This was the primary cause of the regression.

2. yolov9 had no path fix at all — it was listed in the amend formats list but had no callback case, so the YAML was read and rewritten with zero changes.

3. lstrip("..") was misused — str.lstrip strips any combination of the given characters, not a prefix. This silently mangled paths like ../../train/images and .hidden/train/images.

4. test key was always written even when the dataset had no test split, producing a path to a non-existent directory.

Fix

• Removed the ultralytics version gate entirely — paths are now always corrected for yolov8 and yolov9, regardless of whether ultralytics is installed or what version it is.
• Added yolov9 to the relative-path branch (train/images, valid/images, test/images).
• Replaced lstrip with _strip_rel_prefix — a proper helper that iteratively removes ../ and ./ prefixes without mangling legitimate path characters.
• Used self.splits["test"] > 0 to decide whether to emit the test key; it is removed when the split is absent.
• Used os.path.join instead of string concatenation for absolute paths in legacy formats.

Behavior matrix

Format	Path style	Before	After
yolov8	relative	../train/images (broken unless ultralytics==8.0.196)	train/images
yolov9	relative	../train/images (always broken)	train/images
yolov5pytorch	absolute	<location>train/images (lstrip bug)	<location>/train/images
yolov7pytorch	absolute	<location>train/images (lstrip bug)	<location>/train/images
mt-yolov6	absolute	<location>train/images (lstrip bug)	<location>/train/images

Test plan

• 13 new tests in TestReformatYaml covering:
  • yolov8/yolov9 use relative paths (no ../ prefix)
  • yolov8/yolov9 omit test key when no test split exists
  • yolov5pytorch/yolov7pytorch/mt-yolov6 use absolute paths
  • Legacy formats omit test key when no test split exists
  • ../../ double-parent prefix is correctly stripped
  • Non-YOLO formats (e.g. coco) are not modified
  • Works without ultralytics installed
  • Works with any ultralytics version
  • Other YAML fields (nc, names, roboflow) are preserved
• All 804 existing tests still pass
• ruff format --check passes
• ruff check passes
• mypy roboflow passes

Generated with Devin

[davidnichols-ops]

Thanks for the review consideration! Quick summary of what this PR does and why it's a clean merge:

Scope: Single method (__reformat_yaml), 2 files, 1 commit. No unrelated changes.

The fix addresses 4 bugs in one method:

1. Removed the ultralytics==8.0.196 version gate that caused the regression — paths are now always corrected regardless of installed package version
2. Added yolov9 to the relative-path branch (was listed in amend formats but had no callback case)
3. Replaced lstrip("..") with proper prefix stripping (lstrip strips any combination of chars, not a prefix — silently mangled ../../ paths)
4. Only writes test key when self.splits["test"] > 0 (was always written, even for datasets without a test split)

Verification:

• 804 existing tests pass (no regressions)
• 13 new tests covering all 5 YOLO formats, test-split presence/absence, ultralytics independence, edge cases
• ruff format --check, ruff check, and mypy all clean

Happy to address any feedback!