[model cards] Keep evaluation order in training logs if there's multiple evaluators

[tomaarsen]

Hello!

Pull Request overview

• Keep evaluation order in training logs if there's multiple evaluators
• Rename "loss" to Validation Loss if there's only one loss

Details

I move away from the set-based implementation to a list-based one, so I can reuse the original ordering of the evaluators used in a SequentialEvaluator passed to the Trainer. This gives us e.g.:

Instead of e.g.

gooaq-1024-dev_cosine_map@100	gooaq-128-dev_cosine_map@100	gooaq-256-dev_cosine_map@100	gooaq-32-dev_cosine_map@100	gooaq-512-dev_cosine_map@100	gooaq-64-dev_cosine_map@100

Which was ordered based on the evaluation name (which isn't always useful, as can be seen here)

Also, we used to have

Epoch	Step	Training Loss	loss

which just looked a bit odd, "loss" is now formatted as "Validation Loss" if there's only one loss.

• Tom Aarsen