Hi all,

I was waiting for the leadboard to become stable before sharing my notebooks (https://github.com/AMR-KELEG/projects-portfolio/tree/master/%5BZINDI%5DTo%20Vaccinate%20or%20Not%20to%20Vaccinate).

I have fine-tuned BERT large model using the huggingface transformers package and developed all my models using Google Colab.

My top tips for this competition/ general tips are:

• Instead of picking a single prediction label, compute an average values as sum(label_probability * label).

e.g: If the probabilities of the labels (-1, 0, 1) are (0.6, 0.3, 0.1), then the model's prediction will be -0.6 + 0.1 = -0.5

Generally, deep learning models use Softmax layers for classification problems which generate such probabilities as their output.

• I still struggle with reproducing the results of my models so I generally save the model's weights after training.

It would be a pity if you can't get the perfect results that you have achieved on one of your runs.

• I track the scores of my models in a spreadsheet and try to keep the details of my models there as comments to keep track of my results and ideas.