Have dataset items that are somewhat evenly of each type. The LLM learned some of the types fine. However rotated images are causing problems.
The image sizes are between 1 and 10 pixels.
Here the majority of dataset items are rotated images. Since this is what my LLM is struggling with.
Smaller images. Here the image sizes are between 1 and 5 pixels.
This helped a lot on the validation loss.
Main focus is now on… See the full description on the dataset page:
https://huggingface.co/datasets/neoneye/simon-arc-image-v33.
