Hi
@yhaddad and
@stubbi Thanks for your input

Thank you both for your valuable input.
Initially, we trained our model using a subset of the data for both training and testing, and this yielded a promising R² score of around 81. As we gradually increased the dataset size for both training and testing—while keeping the model architecture unchanged—we continued to observe good results. However, we noticed a significant increase in training time, reaching up to 2 hours per epoch. Interestingly, the MSE score decreased significantly during this phase, indicating better learning.
We also experimented with training on the full dataset (around 1,000,000 samples) while using a smaller validation subset. This setup resulted in an R² score of 74, which still demonstrates good performance.
From these observations, we conclude that our model is scalable for larger datasets, even with a relatively small number of qubits. It also shows potential to become more robust when slightly deepened by adding a few more layers. So far, we have intentionally kept the model architecture unchanged to ensure fair evaluation before the challenge deadline.
the thing is that it was not specified in the problem statement that we need to train on the entire dataset for the final metrics on the challenge backend which I believe was the reason for the dimensionality error. With additional compute resources, we could have reduced the training time and likely resolved the final submission issue.
Since the challenge backend doesn’t allow us to explicitly limit the evaluation to a data subset, we are adopting the approach suggested by
@yhaddad:
We will train the model on a small data sample and evaluate it on the entire dataset.
and we are also trying the second method to split the train dataset for our training and validation both . We will update you with the results shortly.