[Metric Request] Fuzzy reordering and Kendall tau distance

Hello! Looking through this amazing paper and blog-post from a couple years back:
https://arxiv.org/pdf/2109.01396.pdf
https://lena-voita.github.io/posts/nmt_training_through_smt_lens.html
https://www.youtube.com/watch?v=LB6USxmxxEQ

It seems as if tracking Encoder/Decoder model training using BLEU or similar metrics only gets you so far - the "fuzzy reordering score" and "Kendall tau" continue to show improvement when other metrics have stagnated (and indeed there are qualitative improvements to the metrics as well).

The closest thing I could find is for the first one, written in good old Python 2:
https://github.com/google/topdown-btg-preordering/blob/master/evaluate_preordering.py

Both metrics seem relatively straightforward (famous last words!).

A huge thank you to HF and the amazing community that is massively democratizing ML!