It depends on positions and time control. On moderately hard 6-men positions, as you see here, about 90% at 0.25s/move and a bit less at 0.10s/move. But in real games things are different. About 70-75%, maybe more, of 6-men positions occurring in normal games are so unbalanced or so balanced, that there is no need of any TBs, the games are pretty decided as Wins or Draws, and Stockfish converts almost all of them to the correct outcome without any TBs. The remaining, maybe 20-25% of 6-men positions occurring in normal games are trickier, and Stockfish by itself (no-TB) might fail to convert a significant part of them, say on average 50%. And here Syzygy-5 starts helping greatly, reducing that to say 10%.Nordlandia wrote:In general, how often do 5-men convert 6-men positions.
How often do stockfish dev equipped with 5-men convert 6-men positions.
That's why it's important to have good test suites. I do have many gradations of hardness 6-men suites, from regular, easy ones, occurring in normal games, to very hard ones for engines (and usually humans).
So, from regular games, 6-men positions with 5-men Syzygy are converted in more than 95% of cases. From my hardest 6-men suites, the conversion rate with 5-men Syzygy is below 60%.
I use the harder ones as magnifying glass for TB implementation, the only snag is that the hardest ones are not very long (not so many positions), and might be preferring only several few types of positions.
If one is using regular openings, say 2moves_v1.epd to test TB implementation, like Fishtest is dumbly doing, not only they are dealing with the easiest, regular 6-men positions (those 95% solved) occurring in games, but 80+% of games don't even enter a phase to rely on TBs at all. Their resolution power is orders of magnitude lower than mine.