rl_config.json

{
  "mode": "sf",
  "n_games": 32,
  "mcts_sims": 200,
  "mcts_batch_size": 192,
  "mcts_c_puct": 2.5,
  "visit_temp": 1.0,
  "ply_cap": 200,
  "sf_elo": 3200,
  "sf_full_strength": true,
  "sf_depth": 8,
  "sf_move_time": 0.05,
  "root_noise_frac": 0.25,
  "record_sf_moves": true,
  "train_epochs": 1,
  "train_batch_size": 192,
  "train_lr": 1e-05,
  "value_weight": 0.5,
  "grad_clip": 0.5,
  "mix_sf_frac": 0.0,
  "sf_shard_dir": null,
  "mix_sf_move_frac": 0.5,
  "dataset_dir": "outputs/rl_selfplay_a40_soft/dataset",
  "iterations": 40,
  "games_per_iter": 32,
  "eval_games": 8,
  "eval_sims": 100,
  "checkpoint": null,
  "prior_checkpoint": null,
  "output_dir": "outputs/rl_selfplay_a40_soft",
  "use_fp16": false,
  "use_bf16": true,
  "num_workers": 1
}