Wrong number of classes in data.yaml

[sannicosan]

Description

After labelling all the images in the dataset, when generating a new dataset version and downloading it, the data.yaml created additional non-existing classes.

Steps to reproduce

1. Label your data
2. Create a new dataset version
3. Check the Modify Classses pre-processing step to make sure you have the right classes
4. Generate the dataset version
5. Download the dataset
6. Check the classes in data.yaml

Expected result

You should only get the classes that you labeled

Actual result

The roboflow creates additional non-existing classes (this also alters the class_id of the real classes)

Note: While labeling, there were some classes that were created by mistake, but they were deleted afterwards since they were never used. This somehow was still picked up by the roboflow.

Evidence

You can clearly see I had 2 classes:

When downloading the dataset, its data.yaml config file looked like this:

This remapped my original class_id's of course.
This was my original data.yaml:

[sannicosan]

Additional info: As a workaround, I tried to explicitly use the 'Modify class' functionality to make sure the classes were correct. I set this step before generating a new dataset version, and still the data.yaml picked up new classes out of nowhere. You can observe this here:

[FrancescoSaverioZuppichini]

Thanks @sannicosan for opening the issue, I've escalated this internally and we will reply soon :)

[sannicosan]

Here I attach another evidence of how the unexisting/unused class appears out of nowhere:
https://www.loom.com/share/4f17714fa0f34bcc9e0cfd9b2808be76

[FrancescoSaverioZuppichini]

hey @sannicosan could you please add help@roboflow.com to your workspace?

[sannicosan]

Done: https://app.roboflow.com/final-project-object-detection-for-instore-inventory-management

[paulguerrie]

Hi @sannicosan! I took a look at your project, and I was able to export a version while dropping the 'typo class'. Not sure why the '/' class didn't show up in your class remapping window.

Checkout version 28 and see if that gets you what you need.

[SalmanSattar24]

Hello,

The same thing is happening to our data set. We have 5 labels for objects we are looking for circles of color: blue, red, purple, orange, yellow. After running through RoboFlow to convert the Pascal VOC labels to .txt for yoloV8, the data set .txt labels had 10 classes. It created a red-circle, blue-circle, ... etc. The new classes that appear out of nowhere are being assigned to some of the classes have labeled. There is an option to delete classes after the creation of all the .txt files. Does this delete all the false class created to replace them back to what they need to be. Meaning class "blue-circle" will revert back to "blue"