Build a neural network model for imbalanced data and few-shot learning

I need a regression model that should support any imbalanced dataset that have a single continuous target and should be based on neural network architecture that allows the model to learn easily from imbalanced datasets. The model should take raw samples without preprocessing techniques such as: oversampling, undersampling, smote and etc. The model should be capable of few-shot learning, because some datasets are very small, some datasets have only 2000 samples in total and cannot have more, so the model should work well even with datasets which are so small. It is important to not overfit or underfit the model, and build the model with optimal number of trainable parameters, because a optimal means as few parameters as possible to build a small model that is capable to generalise very well. The model evaluation metric should have PRC higher than 95 , F-score > 0.95 , AUCPRC > 95. The model must have as few trainable parameters as possible to keep the model small, and number of model's parameters must not exceed the size of unique samples in the dataset. The model must find common patterns instead of memorizing each sample. I need an experienced freelancer who knows his job very well and who is not wasting his time and mine on simple things saying it is very hard. Usually unexperienced freelancers ask more and provide less value. If you have past experience then this should be something similar and straight forward project for you, otherwise don't apply because I will not pay for time you take to study new things and provide me with poor results. I'm expecting: - delivery deadline: with 1-3 days - budget: $120 If you meet my expectation than more tasks in ML will come, because I'm trying to find a good candidate. Please write the word "Understood" in your bid as first word to check that you read the project description. Test the model on following datasets: - [login to view URL] - Attachment_1