README.md

metadata

dataset_info:
  - config_name: CHUNK_0
    features:
      - name: obs
        sequence:
          sequence:
            sequence: uint8
      - name: target
        sequence: int16
    splits:
      - name: train
        num_bytes: 17589740814
        num_examples: 108578647
      - name: test
        num_bytes: 9883876284
        num_examples: 61011582
    download_size: 2415896482
    dataset_size: 27473617098
  - config_name: SHUFFLED_CONCAT
    features:
      - name: obs
        dtype:
          array3_d:
            shape:
              - 2
              - 6
              - 7
            dtype: float32
      - name: target
        sequence: float32
        length: 7
    splits:
      - name: train
        num_bytes: 207168057840
        num_examples: 488603910
      - name: validation
        num_bytes: 4603734760
        num_examples: 10857865
    download_size: 8213390894
    dataset_size: 211771792600
  - config_name: SHUFFLED_CONCAT_VALIDATION
    features:
      - name: obs
        dtype:
          array3_d:
            shape:
              - 2
              - 6
              - 7
            dtype: float32
      - name: target
        sequence: float32
        length: 7
    splits:
      - name: validation
        num_bytes: 4603734760
        num_examples: 10857865
    download_size: 178569088
    dataset_size: 4603734760
  - config_name: TRAIN_ONLY
    features:
      - name: obs
        dtype:
          array3_d:
            shape:
              - 2
              - 6
              - 7
            dtype: float32
      - name: target
        sequence: float32
        length: 7
    splits:
      - name: train
        num_bytes: 46037346328
        num_examples: 108578647
    download_size: 1549311584
    dataset_size: 46037346328
  - config_name: default
    features:
      - name: obs
        dtype:
          array3_d:
            shape:
              - 2
              - 6
              - 7
            dtype: float32
      - name: target
        sequence: float32
        length: 7
    splits:
      - name: test
        num_bytes: 25868910768
        num_examples: 61011582
      - name: train
        num_bytes: 46037346328
        num_examples: 108578647
    download_size: 2419867693
    dataset_size: 71906257096
configs:
  - config_name: CHUNK_0
    data_files:
      - split: train
        path: CHUNK_0/train-*
      - split: test
        path: CHUNK_0/test-*
  - config_name: SHUFFLED_CONCAT
    data_files:
      - split: train
        path: SHUFFLED_CONCAT/train-*
      - split: validation
        path: SHUFFLED_CONCAT/validation-*
  - config_name: SHUFFLED_CONCAT_VALIDATION
    data_files:
      - split: validation
        path: SHUFFLED_CONCAT_VALIDATION/validation-*
  - config_name: TRAIN_ONLY
    data_files:
      - split: train
        path: TRAIN_ONLY/train-*
  - config_name: default
    data_files:
      - split: test
        path: data/test-*
      - split: train
        path: data/train-*
license: mit
task_categories:
  - reinforcement-learning
size_categories:
  - 100M<n<1B

Connect 4 Solver Outputs

About 100M + 60M observations and targets generated from selfplay with a solver [https://github.com/PascalPons/connect4] with temperature. Observations from different depths are roughly uniformly distributed, altough later positions are reached less frequently. As a consequence early positions are duplicated and there is a small overlap between the train and test split (less than 3%).

Observations are of shape (2,6,7) with binary (0 or 255) data. The first channel represents the stones placed by the current player while the second channel represent the stones played by the opoonent. The targets are the scores returned by the solver for placing the next stone in earch column. E.g. 1 / -1 means winning / losing on the last possible move, 2 / -2 means winning / losing one move before etc.