Training of GPT-Neo or DistilBERT for automatic marking

I am building AI software for education that will automatically mark student responses to questions. I want to analyse how effectively a small pre-trained model such as GPT-Neo or DistilBERT could be trained using 200 responses for this task, and whether it could compete with GPT 4 on that specific domain. I will provide: The text of a maths question, and a mark scheme for that question 200 different student responses to the question and the associated marks I need you to train a suitable small model and deliver your model in a form that I can use it to test the quality of marking in other unseen responses. With your bid please confirm: - which model you would choose for this task - how long it would take you to do the training - what format the model would be provided to me so that I can test it for future training Thanks very much! Rob