strategic_game_chess

# Chess

> Recent advancements in artificial intelligence (AI) underscore the progress of reasoning and planning shown by recent generalist machine learning (ML) models. The progress can be boosted by datasets that can further boost these generic capabilities when used for training foundation models of various kind. This research initiative has generated extensive synthetic datasets from complex games — chess, Rubik's Cube, and mazes — to study facilitation and the advancement of these critical generic skills in AI models.

This dataset contains 3.2 billion games, equating to approximately 608 billion individual moves.
it is generated through self-play by Stockfish engine using Fugaku and we add initial moves to expand its diversity.

Each game has three columns: 'Moves', 'Termination' and 'Result',
- 'Move': recorded chess moves of the whole game.
- 'Termination': include CHECKMATE, INSUFFICIENT_MATERIAL, ... etc.
- Please check this for detail information

https://python-chess.readthedocs.io/en/latest/core.html#chess.Outcome.termination
- 'Result': result of this game, 1-0, 1/2-1/2, 0-1.

### Call for Collaboration

We invite interested researchers and ML practitioners to explore these datasets' potential. Whether training GPT models from scratch or fine-tuning pre-existing models, we encourage the exploration of various pre-training and fine-tuning strategies using these game-based datasets standalone or as enhancement of other already composed large-scale data.

Our team is prepared to assist in securing necessary GPU resources for these explorations. We are particularly interested in collaborators eager to pre-train models of small to medium scale on our game data, subsequently transition to standard text-based training, and then perform comparative analyses against models of similar architecture trained exclusively on text data.

Conclusively, this initiative marks a significant stride toward intricate problem-solving and strategic planning in AI, extending an open invitation to the research community for collaborative advancement in this domain.

# 国际象棋（Chess）数据集

> 近年来人工智能（AI）领域的进展，凸显了当前通用型机器学习（ML）模型在推理与规划能力上的突破。若将此类数据集用于训练各类基础模型，可进一步强化这些通用能力，从而加速上述进展。本研究项目从国际象棋、魔方（Rubik's Cube）与迷宫等复杂博弈场景中生成了大规模合成数据集，用于探究如何助力并提升AI模型的上述关键通用技能。

本数据集共包含32亿局对局，总计约6080亿步棋步。该数据集通过在富岳超级计算机（Fugaku）上运行Stockfish引擎进行自我对弈生成，并通过补充初始棋步以提升数据集的多样性。

每局对局包含三列数据：'Moves'、'Termination'与'Result'：
- 'Moves'：记录完整对局的所有棋步。
- 'Termination'：对局终止原因，包含将死（CHECKMATE）、子力不足（INSUFFICIENT_MATERIAL）等类型，详细说明可参考：
https://python-chess.readthedocs.io/en/latest/core.html#chess.Outcome.termination
- 'Result'：对局最终结果，取值为1-0、1/2-1/2或0-1。

### 合作招募

我们诚邀感兴趣的研究人员与机器学习从业者挖掘本数据集的应用潜力。无论您计划从零开始训练GPT模型，还是对已有模型进行微调，我们都鼓励您探索基于此类博弈数据集的各类预训练与微调策略——既可单独使用该数据集，也可将其作为现有大规模数据集的增强补充。

本研究团队可协助协调相关探索所需的GPU算力资源。我们尤其欢迎有意向使用本博弈数据集对中小型规模模型进行预训练、随后转向标准文本训练，并最终与仅基于文本数据训练的同架构模型开展对比分析的合作者。

综上，本项目在AI复杂问题求解与战略规划领域迈出了重要一步，我们谨向全球研究社区开放合作邀请，共同推动该领域的发展。