yethdev/2048-gameplay-dataset
收藏Hugging Face2026-04-09 更新2026-04-12 收录
下载链接:
https://hf-mirror.com/datasets/yethdev/2048-gameplay-dataset
下载链接
链接失效反馈官方服务:
资源简介:
---
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:
```json
{
"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
许可协议:其他
许可协议名称:MIT署名协议(mit-attribution)
任务类别:
- 强化学习(reinforcement-learning)
标签:
- "2048"
- 游戏AI(game-ai)
- 监督学习(supervised-learning)
- 模仿学习(imitation-learning)
样本量范围:10000 < 样本量 < 100000
# 2048游戏专家对局数据集
本数据集包含由专家级N元组网络(N-Tuple Network)AI智能体(AI Agent)游玩2048游戏时生成的状态-动作配对数据,可用于2048游戏智能体的模仿学习与监督训练。
## 统计数据
- 原始对局数:10000局
- 过滤后有效对局数:9000局
- 总动作步数:54010983步
- 平均单局得分:143847
- 胜率(达成2048及以上数值方块):100%(已剔除未达成目标的失败对局)
- 得分下限:62152(已剔除得分排名后10%的对局)
## 文件说明
- `train.jsonl`:训练集,包含8100局对局,共计48592790步动作
- `val.jsonl`:验证集,包含900局对局,共计5418193步动作
- `games.jsonl`:原始未过滤数据集,包含10000局对局
## 过滤规则
为提升监督训练的质量,我们对原始数据集进行了如下过滤处理:
1. 剔除未达成2048数值方块的失败对局(包含绝境收尾动作的对局)
2. 剔除得分排名后10%的对局(包含因运气不佳导致棋盘杂乱的对局)
3. 按照90:10的比例划分为训练集与验证集(每10局保留1局作为验证集)
## 数据格式
数据集采用JSONL格式,每行代表一局完整对局。每局对局对应的JSON对象结构如下:
json
{
"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`:以行优先顺序存储的16个方块数值(0代表空槽,2/4/8等代表对应数值的方块)
- `action`:动作编码,0对应上移,1对应右移,2对应下移,3对应左移
- `reward`:本次移动合并方块所获得的得分
提供机构:
yethdev



