When training ranking models, should we use explicit negative samples or negative samples sampled from the batch, or provided by the retrieval model?

[viswa-nvidia]