Training log details

[dandelin]

Hi! Thank you so much for sharing your work :)

Actually, I've been re-implementing MERU and found that the learnable curvature quickly drops to its lower bound of 0.1.
Also, the entailment loss was near zero during the training. (Though zero-shot performances on the classification and retrieval were fine.)

As I'm unsure if this is normal, before scrutinizing your code and re-running the training session with this repo, I would be grateful if you could share the training logs.

Thanks!