Adam9044/machiavelli
收藏资源简介:
该数据集是“The MACHIAVELLI Benchmark”的派生版本,旨在将原始基于强化学习/代理的文本冒险游戏数据集转换为适合大型语言模型使用的格式。数据集用于评估LLM的道德性,无需微调,包含来自标注文本冒险游戏的轨迹节点、选择和标签。数据特征包括游戏标题、玩家角色、简短摘要、观察文本、早期和近期历史序列、行动序列,以及每个选择对应的道德标签(如合作、诚实、多种道德维度)和奖励。数据集分为训练集和测试集,适用于自然语言处理中的道德推理和决策分析任务。
This dataset is a fork of The MACHIAVELLI Benchmark, designed to convert the original RL/agent-based text adventure dataset into a format suitable for large language models. It is used for evaluating the morality of LLMs without fine-tuning, containing trajectory nodes, choices, and labels from annotated text adventure games. Features include game title, player character, short summary, observation text, early and recent history sequences, action sequences, and moral labels (e.g., cooperation, honesty, various moral dimensions) and rewards for each choice. The dataset is split into training and test sets, suitable for moral reasoning and decision analysis tasks in natural language processing.




