module

Experion::GenData

Inherits Experion::Moves

Instance methods

run(games : Int32, movetime_ms : Int64, path : String, threads : Int32 = 1) : Nil
Source
run_classical(positions : Int32, path : String, threads : Int32 = 1) : Nil

Classical-distillation mode: random-walk positions labeled with the current classical evaluation (white POV centipawns). No searching — produces millions of samples quickly so the NNUE can learn material, PSTs and basic terms before result-based fine-tuning.

Source
run_scored(positions : Int32, depth : Int32, path : String, threads : Int32 = 1) : Nil

Scored-random mode: random-walk positions labeled by a FIXED-DEPTH search score. Labels carry tactical information beyond the static evaluator while remaining cheap enough to generate at scale.

Source
run_selfplay(games : Int32, depth : Int32, path : String, threads : Int32 = 1) : Nil

Depth-limited self-play: real games (not random walks) labeled with the engine's own search score and the eventual game result. Opening diversity comes from a few random plies plus occasional non-best moves, not from a second engine's eval.

Source