Add on-the-fly sampling

[lebionick]

As far as I understand, right now text pairs are picked directly using Datapack samples. However it is almost impossible to mine "hard" negative samples before training. Just adding into class 0/ rank 0.0 random pairs which are not corresponding is very wasteful and inefficient.

[faneshion]

Hi @lebionick , do you mean the sampling strategy should be dynamically adjusted according to the performance after each epoch?

[lebionick]

@faneshion exactly