Please see #29 .

I have actually gotten some interesting results that differ from the binary W/L ratings, by passing fractional numbers as ranks. The results are not very predictive, but then again, the TrueSkill/OpenSkill model so far appears to be less predictive than Glicko or Elo with non-fractional ranks for my use case in general, so I'm not sure whether the fractional ranks are doing what I would like to use them for.

As far as data goes, the example given in the linked thread should work well. I understand the response that goals in an 11 v 11 game may not be a good metric because of positional roles, but in a team vs. team context, final score should be more predictive than simple win/loss (eg. Team A beating Team B 7-0 is more predictive of strength than Team A beating Team B 2-1).

In a multiplayer FFA context, kills/deaths should be more predictive than ordinal rank. For example, Player 1 with 11 kills and 0 deaths, Player 2 with 1 kill and 6 deaths, Player 3 with 1 kill and 7 deaths, would rank Player 1 way ahead of Player 2 and 3, whereas giving their ranks as 1, 2 and 3 would not really give a good picture of relative skill.