game: connect4
algo: efficientzero
wandb_project: crpt
wandb_enabled: true
checkpoints:
  save: true
  dir: null
  every_env_steps: 10000
  best_ckpt_strategy: raw
  best_ckpt_ema_alpha: 0.3
  best_ckpt_min_episodes: 20
  max_checkpoints_to_keep: 3
  load_from: /home/molfetta/combinatorial_reasoning_post_training/models/main5_hf_bot_mode_200k_collector_tuned_20260522/connect4/attempt-01_260522_142412/ckpt/ckpt_last.pth.tar
evaluation:
  opponent_type: env_bot
  fixed_bot_evaluator:
    type: arena
    seat_swap: true
  every_train_iters: null
  every_env_steps: 10000
  previous_best_checkpoint:
    path: null
    selector: best
    update_policy: on_new_best
    num_simulations: null
    n_evaluator_episode: null
    evaluator_env_num: null
    promotion_threshold: 0.0
    fallback_to_env_bot: false
  runtime_battle_mode: eval_mode
  opponent_regime: meaningful_fixed_bot
  opponent_impl: handwritten:lightzero_rule
  primary_metric: win_rate_vs_fixed_bot
  bot_strength_tier: moderate
  bot_deterministic: false
  requires_paired_audit: false
  meaningful_fixed_bot: true
env:
  collector_env_num: 32
  evaluator_env_num: 20
  n_evaluator_episode: 20
  battle_mode: play_with_bot_mode
  bot_action_type: rule
  prob_random_action_in_bot: 0.0
  battle_mode_in_simulation_env: self_play_mode
  extra_config:
    collector_bot_mode_seat_swap: true
  stop_value: 2
defaults:
  seed: 0
  max_env_step: 1000000
  num_simulations: 50
  batch_size: 256
  learning_rate: 0.003
  piecewise_decay_lr_scheduler: false
  update_per_collect: 25
  replay_buffer_size: 100000
  discount_factor: 1
  game_segment_length: 21
  td_steps: 21
  reanalyze_ratio: 0.0
  num_unroll_steps: 5
model:
  num_res_blocks: 1
  num_channels: 64