Processing Dataset of 30Mb makes Google Colab's 13Gb RAM run out of memory

[Koenlaermans]

I'm trying to fine tune a deepset QA reader from hugging face on the CUAD dataset (a QA dataset of about 20k questions). I'm only using 3500 questions right now.
With only these 3 lines of code I'm trying to fine tune the reader on this dataset. The json file in squad format is about 27.5Mb. Still the 12 Gb System-RAM of my google colab runtime is running out of memory while preprocessing the dataset (not even training).
The problem is not the GPU but System RAM. I don't know why this is or how I would fix this. It doesn't seem logical.
Link to the dataset: CUAD Dataset 3500 questions

Here is the code.

reader = FARMReader("deepset/minilm-uncased-squad2", use_gpu=True)
data_dir = "drive/MyDrive/Colab Notebooks"
reader.train(data_dir=data_dir, train_filename="CUAD.json", use_gpu=True, n_epochs=2, save_dir="drive/MyDrive/Colab Notebooks/TrainedMiniOnCUAD2")