Skip to content

KataGomo Y, Obtuse-Y, Bent-Y 20260716

Latest

Choose a tag to compare

@hexanna1 hexanna1 released this 17 Jul 05:16

b10c256nbt-fson-rvgl-bnh architecture, 7.1 million parameter model. Bent-Y uses a smaller b5c192nbt-fson-rvgl-bnh architecture with 2.1 million parameters.

The y24 net plays Y. It was trained progressively on board sizes up to 9, then <=15, <=21, and briefly on <=24.

The oy21 net plays Obtuse-Y. It was trained progressively on board sizes up to 6, then <=9, and briefly on <=11. A size-11 hexhex board is embedded in a 21×21 Hex grid, hence the name.

The by9 net plays Bent-Y. It was trained on board size 5, and then a mixture of sizes 5 and 9.

The Y weights and config can be used with a compiled engine in the Hexata GUI.