buzzwords-agent-trace
收藏资源简介:
该数据集名为“Buzzwords & Misdemeanors — agent traces”,包含一系列JSON文件,每个文件记录了一局完整的“Buzzwords & Misdemeanors”游戏(一个法庭模拟游戏)。数据通过基于代理的交互系统生成:一个“Game Master”(使用MiniCPM5-1B基础模型和蒸馏的director LoRA,并受GBNF约束)按节拍指导听证会,基于代码中采样的真相;确定性守卫强制执行法庭序列不变性;多个“actors”(使用相同基础模型和每个风格的LoRA)负责交付台词。每个轨迹详细记录了每个节拍中Game Master的原始结构化决策,包括说话者、节拍类型、事实索引线索通道、强度、舞台指导和总结,以及守卫是否重新映射了决策、生成的台词、隐藏的真相和评分器校准探针。所有推理均在CPU上使用llama.cpp运行,依赖于一个约10亿参数的基础模型和小型适配器。该数据集适用于研究代理轨迹、法庭游戏模拟、语言模型在结构化环境中的应用,以及多智能体交互系统的评估。
The dataset is named Buzzwords & Misdemeanors — agent traces and consists of a series of JSON files, each recording a complete game session of Buzzwords & Misdemeanors (a courtroom simulation game). The data is generated through an agent-based interaction system: a Game Master (using the MiniCPM5-1B base model with a distilled director LoRA and constrained by GBNF) guides the hearing in beats based on truth sampled from code; a deterministic guard enforces courtroom sequence invariance; multiple actors (using the same base model with LoRAs for each style) deliver lines. Each trace details the Game Masters original structured decisions per beat, including speaker, beat type, fact index clue channel, intensity, stage direction, and summary, as well as whether the guard remapped the decision, generated lines, hidden truth, and scorer calibration probes. All inference is run on CPU using llama.cpp, relying on a base model with approximately 1 billion parameters and small adapters. This dataset is suitable for researching agent trajectories, courtroom game simulation, language model applications in structured environments, and evaluation of multi-agent interaction systems.
数据集概述
- 数据集名称:Buzzwords & Misdemeanors — agent traces
- 许可证:Apache-2.0
- 标签:agent-trace, llama-cpp, courtroom-game
数据集内容
该数据集包含多轮法庭场景游戏“Buzzwords & Misdemeanors”的完整对局记录。每一份JSON文件对应一局完整的游戏过程。
数据构成
每个对局记录包含以下结构化信息:
- 游戏大师(Game Master):基于MiniCPM5-1B模型与经过蒸馏的“导演”LoRA适配器,受GBNF约束,按节拍指导听证过程。真相通过代码采样生成。
- 确定性守卫(Deterministic Guards):确保法庭流程顺序的不变性。
- 演员(Actors):使用相同的基座模型搭配每种风格对应的LoRA适配器来生成台词。
- 每节拍记录:
- 发言者
- 节拍类型
- 事实索引/线索通道
- 强度
- 舞台指示
- 总结
- 守卫是否对节拍进行了重映射
- 最终生成的台词
- 隐藏的真相
- 评分者校准探查
配置与文件
- 配置名称:
agent_traces_actor_line - 数据文件路径:
traces/* - 分割:仅包含训练集(
train)
技术细节
所有推理均使用llama.cpp在CPU上运行,采用约1B参数的基座模型搭配小型适配器。




