1. Take Frank's pgn, this is at tc 10min/40moves with good hardware.brtzsnr wrote: Are there any other ideas I could try to pick better positions for tuning?
2. Separate it by eco, try to balance number of games in every eco.
3. Extract positions with move score <= 6 pawns and score >= -0.5, side to move has won. (Be aware the move score in that pgn is white POV).
4. Extract positions from drawn games up to 4 to 6 pos per game. Just to counter that not all 2-bishop advantage is a win for example.
5. If the engine under tuning is using qsearch score and is not generating check evasion in qsearch, then extract only positions where the side to move is not in check.
Take other sources one that is from 2 move, 3 move, 4 move, 5 move, 6 move, 7 move, 8 move, 9 move, 10 move opening start pos. Frank's pgn is probably in the range 10 to 14 book moves. This way you have a good coverage of positions.
From experience with the Texel method, move eval method and GA Stefano method, these positions give improvement in Deuterium's performance.