Upload latest retained checkpoint envstep_300000.pth.tar

metadata/metadata.yaml

@@ -0,0 +1,215 @@
+metadata_version: 2
+saved_at: '2026-05-21T05:09:21.229267+00:00'
+entrypoint: src.train_game
+game: three_six_nine
+algo: efficientzero
+seed: 0
+max_env_step: 300000
+exp_name: /workspace/combinatorial_reasoning_post_training/models/simplified5_two_machine_continuation_20260520/shard-0/three_six_nine/attempt-03_260520_145623
+checkpoint_dir: /workspace/combinatorial_reasoning_post_training/models/simplified5_two_machine_continuation_20260520/shard-0/three_six_nine/attempt-03_260520_145623/ckpt
+config_path: /workspace/combinatorial_reasoning_post_training/data/supported_training/runs/simplified5_two_machine_continuation_20260520/shard-0/configs/three_six_nine.yaml
+cli_args:
+  game: null
+  config: /workspace/combinatorial_reasoning_post_training/data/supported_training/runs/simplified5_two_machine_continuation_20260520/shard-0/configs/three_six_nine.yaml
+  list_games: false
+  algo: null
+  num_simulations: null
+  batch_size: null
+  update_per_collect: null
+  lr: null
+  seed: null
+  max_env_step: null
+  num_res_blocks: null
+  num_channels: null
+  replay_buffer_size: null
+  num_unroll_steps: null
+  reanalyze_ratio: null
+  battle_mode: null
+  no_wandb: false
+  cpu: false
+  torch_num_threads: 4
+  torch_num_interop_threads: 1
+  no_checkpoints: false
+  ckpt_dir: /workspace/combinatorial_reasoning_post_training/models/simplified5_two_machine_continuation_20260520/shard-0/three_six_nine/attempt-03
+  model_path: null
+  ckpt_every_env_steps: null
+  max_checkpoints_to_keep: null
+  best_ckpt_strategy: null
+  best_ckpt_ema_alpha: null
+  best_ckpt_min_episodes: null
+  eval_every_train_iters: null
+  eval_every_env_steps: null
+  eval_every_env_step_ratio: null
+  eval_opponent_type: null
+  eval_opponent_checkpoint_path: null
+  eval_opponent_checkpoint_selector: null
+  eval_opponent_update_policy: null
+  eval_opponent_num_simulations: null
+  eval_opponent_episodes: null
+  eval_opponent_env_num: null
+  eval_opponent_promotion_threshold: null
+  eval_opponent_fallback_to_env_bot: false
+  wandb_project: crpt-simplified5-corrected-quality
+  wandb_run_name: simplified5_two_machine_continuation_20260520__shard-0__three_six_nine__a03
+  wandb_group: simplified5_two_machine_continuation_20260520
+  wandb_run_id: crpt-3367d2521da9e2789a06
+  wandb_dir: /workspace/combinatorial_reasoning_post_training/data/supported_training/runs/simplified5_two_machine_continuation_20260520/shard-0/wandb
+source_config:
+  game: three_six_nine
+  algo: efficientzero
+  wandb_project: crpt
+  wandb_enabled: true
+  checkpoints:
+    save: true
+    every_env_steps: 5000
+    max_checkpoints_to_keep: 10
+    load_from: /workspace/combinatorial_reasoning_post_training/models/simplified5_two_machine_continuation_20260520/shard-0/three_six_nine/attempt-01_260520_133229/ckpt/envstep_10000.pth.tar
+  evaluation:
+    every_train_iters: null
+    runtime_battle_mode: eval_mode
+    opponent_regime: meaningful_fixed_bot
+    opponent_impl: line_goal
+    primary_metric: win_rate_vs_fixed_bot
+    bot_strength_tier: moderate
+    bot_deterministic: true
+    requires_paired_audit: false
+    meaningful_fixed_bot: true
+    opponent_type: env_bot_and_previous_best_checkpoint
+    previous_best_checkpoint:
+      selector: best
+      update_policy: on_new_best
+      promotion_threshold: 0.0
+      fallback_to_env_bot: false
+      num_simulations: 10
+      evaluator_env_num: 10
+      n_evaluator_episode: 10
+    every_env_steps: 5000
+  env:
+    battle_mode: self_play_mode
+    battle_mode_in_simulation_env: self_play_mode
+    bot_action_type: rule
+    prob_random_action_in_bot: 0.0
+    collector_env_num: 4
+    evaluator_env_num: 10
+    n_evaluator_episode: 10
+    extra_config:
+      eval_random_start_enabled: true
+      eval_random_start_min_moves: 0
+      eval_random_start_max_moves: 20
+      eval_random_start_even_ply: true
+    stop_value: 2
+  defaults:
+    seed: 0
+    num_simulations: 50
+    batch_size: 256
+    update_per_collect: 50
+    learning_rate: 0.003
+    replay_buffer_size: 100000
+    discount_factor: 1
+    game_segment_length: 81
+    td_steps: 81
+    reanalyze_ratio: 0.0
+    num_unroll_steps: 5
+    piecewise_decay_lr_scheduler: false
+    max_env_step: 300000
+  model:
+    num_res_blocks: 1
+    num_channels: 64
+model_info:
+  policy_type: efficientzero
+  model_class: EfficientZeroModel
+  trainable_params: 12626451
+  model_config:
+    observation_shape:
+    - 3
+    - 9
+    - 9
+    action_space_size: 81
+    image_channel: 3
+    num_res_blocks: 1
+    num_channels: 64
+  num_res_blocks: 1
+  num_channels: 64
+  observation_shape:
+  - 3
+  - 9
+  - 9
+  action_space_size: 81
+  image_channel: 3
+resolved_main_config:
+  exp_name: /workspace/combinatorial_reasoning_post_training/models/simplified5_two_machine_continuation_20260520/shard-0/three_six_nine/attempt-03_260520_145623
+  env:
+    battle_mode: self_play_mode
+    bot_action_type: rule
+    channel_last: false
+    collector_env_num: 4
+    evaluator_env_num: 10
+    n_evaluator_episode: 10
+    manager:
+      shared_memory: false
+    battle_mode_in_simulation_env: self_play_mode
+    prob_random_action_in_bot: 0.0
+    eval_random_start_enabled: true
+    eval_random_start_min_moves: 0
+    eval_random_start_max_moves: 20
+    eval_random_start_even_ply: true
+  policy:
+    model:
+      observation_shape:
+      - 3
+      - 9
+      - 9
+      action_space_size: 81
+      image_channel: 3
+      num_res_blocks: 1
+      num_channels: 64
+    model_path: /workspace/combinatorial_reasoning_post_training/models/simplified5_two_machine_continuation_20260520/shard-0/three_six_nine/attempt-01_260520_133229/ckpt/envstep_10000.pth.tar
+    cuda: true
+    env_type: board_games
+    action_type: varied_action_space
+    game_segment_length: 81
+    update_per_collect: 50
+    batch_size: 256
+    optim_type: Adam
+    learning_rate: 0.003
+    grad_clip_value: 0.5
+    num_simulations: 50
+    reanalyze_ratio: 0.0
+    td_steps: 81
+    discount_factor: 1
+    n_episode: 4
+    eval_freq: 300001
+    replay_buffer_size: 100000
+    collector_env_num: 4
+    evaluator_env_num: 10
+    use_wandb: true
+    best_ckpt_strategy: raw
+    best_ckpt_ema_alpha: 0.3
+    best_ckpt_min_episodes: 20
+    num_unroll_steps: 5
+    battle_mode: self_play_mode
+    battle_mode_in_simulation_env: self_play_mode
+    piecewise_decay_lr_scheduler: false
+    mcts_ctree: true
+    eval_opponent_type: env_bot_and_previous_best_checkpoint
+    previous_best_checkpoint:
+      path: null
+      selector: best
+      update_policy: on_new_best
+      num_simulations: 10
+      n_evaluator_episode: 10
+      evaluator_env_num: 10
+      promotion_threshold: 0.0
+      fallback_to_env_bot: false
+    device: cuda
+resolved_create_config:
+  env:
+    type: three_six_nine
+    import_names:
+    - custom_games.three_six_nine.envs.three_six_nine_env
+  env_manager:
+    type: subprocess
+  policy:
+    type: efficientzero
+    import_names:
+    - lzero.policy.efficientzero