Stockfish don't need tablebases. Is reason this?

[Jouni]

In training nn-335a9b2d8a80.nnue there was "rescoring with syzygy 6-piece tablebases and ~600 GB of 7-piece tablebases". So nnue contains TB data already

[Jouni]

I made test playing SF15.1 against classic. Endgame positions and 20 plies search. As usually nnue version was 100 ELO stronger and evaluation was usually much bigger than classic. Examples:

[fen]8/8/8/1p4k1/6p1/4K3/8/3bB3 b - - 0 20[/fen]

classic -1,52 nnue -5,30

[fen]8/7R/4k2P/8/4PK2/8/8/6r1 b - - 0 39[/fen]

classic 1,52 nnue 4,97

[dkappe]

In training nn-335a9b2d8a80.nnue there was "rescoring with syzygy 6-piece tablebases and ~600 GB of 7-piece tablebases". So nnue contains TB data already

The lc0 data has been rescored with TB’s for quite some time independent of Stockfish. Thats because temperature is used to introduce variety while likely throwing off the final result as a side effect. Rescoring with TB’s is a way of correcting for this close to the end.

The pre-rescoring lc0 nets were still pretty good (as was AlphaZero, which didn’t use rescoring).