I think I tried back then but I wasn't getting different scores so I left this as is. I'd keep it as is for now and if we'll do another training on a new dataset I'll get back to this.
Originally posted by @edoardolegnaro in #175 (comment)
I think this will lead to data leakage maybe avoid in future?