Bots are in training dataset

[Tsardoz]

This looks great! I have been looking for something like this that includes human ability in a single model.
Unfortunately there is one major issue with this I am not sure can be fixed.
Previously I have downloaded some of these Lichess datasets.
Based on the supposition that users often use the same accouint name between chess.com and lichess, I extracted all the users, submitted them one by one to chess.com's api, which reports back with a fair play violation if they have banned that user.
Up to 20% of users in recent lichess datasets who have the same name in chess.com have been banned by chess.com.
So your training data will be corrupted by chess AI bots of much higher rating.
I do not know of a way of filtering these out. Maybe you can think of one?

[dav1312]

Took a random 20k player sample from 2M unique players from the March database
52% of them had a matching chesscom account
2.6% of them had a closed account on chesscom

Total players checked: 20000
Total players found:   10521

Player Status Counts:
  - basic                       : 9408
  - premium                     : 386
  - closed:fair_play_violations : 272
  - closed                      : 362
  - closed:abuse                : 93