README.md

metadata

license: other
license_name: mit-attribution
task_categories:
  - reinforcement-learning
tags:
  - '2048'
  - game-ai
  - supervised-learning
  - imitation-learning
size_categories:
  - 10K<n<100K

2048 Expert Gameplay Dataset

State-action pairs from an expert N-Tuple Network agent playing 2048. Can be used for imitation learning / supervised training of 2048 agents.

Stats

• Source games: 10,000
• Games after filtering: 9,000
• Total moves: 54,010,983
• Average score: 143,847
• Win rate (>= 2048): 100% (losing games removed)
• Score floor: 62,152 (bottom 10% removed)

Files

• train.jsonl - 8,100 games (48,592,790 moves) for training
• val.jsonl - 900 games (5,418,193 moves) for validation
• games.jsonl - original unfiltered dataset (10,000 games)

Filtering

The raw dataset was filtered to improve supervised learning quality:

1. Removed games that did not reach the 2048 tile (losing games with desperate end-moves)
2. Removed bottom 10% by score (bad-luck games with messy board patterns)
3. Split 90/10 into train/validation (every 10th kept game to validation)

Format

JSONL format, one game per line. Each game object:

{
  "game_id": 0,
  "score": 142056,
  "max_tile": 8192,
  "num_moves": 2341,
  "moves": [
    {
      "board": [0, 0, 0, 2, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 2, 0],
      "action": 0,
      "action_name": "up",
      "reward": 0,
      "move_number": 0
    }
  ]
}

• board: 16 tile values in row-major order (0=empty, 2/4/8/...)
• action: 0=up, 1=right, 2=down, 3=left
• reward: points from merges on this move