A clear and concise description of what the problem is. Ex. I'm always frustrated when [...]

The model learns more when it is punished more. In Collie, we often use adaptive implicit loss functions that sample multiple negative items at a time to allow for a greater chance to punish the model.

Setting num_negative_samples too low means the model learns slower over time, but setting num_negative_samples might mean the model will be too punished to the point where it does not learn at all.

A clear and concise description of what you want to happen.

Ideally, we could have an adaptive negative sampling strategy that starts with a small num_negative_samples and, as training progresses, will increase the num_negative_samples so the model is given greater chance of messing up as it learns more and more.

A clear and concise description of any alternative solutions or features you've considered.

We could just train for a few epochs, do something like:

# train a few epochs

model.train_loader.num_negative_samples += 5
model.val_loader.num_negative_samples += 5

# continue training

and we would have the same outcome, but it would be nice if this were built into Collie somehow via a tunable hyperparameter!

Add any other context or information about the feature request here.

This would be cool :)