There is pr1682 ("Action replay" - <https://github.com/LeelaChessZero/lc0/pull/1682>) that does selfplay and generates training data from the pgn as if it was played by lc0. But it takes time.