Use stratification in multilabel tasks.

[x-tabdeveloping]

Currently there are multiple multilabel tasks that use dummy or no subsampling leading to potentially unrealistic results.
We will soon have a PR merged that introduces multilabel stratification #760. And it would be very useful to convert all tasks to using this.

These are:

• Hierarchical clustering tasks:
  • SNL
  • VG
  • ArXiv
  • Blurbs
• Multilabel classification tasks:
  • MalteseNews
  • MultiEURLEX
  • BrazilianToxicTweets

[x-tabdeveloping]

@KennethEnevoldsen @imenelydiaker @orionw @isaac-chung @Muennighoff Is this something we want to get done quickly before running everything or should we just roll with things as they are?

[KennethEnevoldsen]

The current subsampling is just random. I don't believe it is a reason to convert them (at least for hierarchical), we can do it for new tasks.

[x-tabdeveloping]

Alright, I will close this then.