build-small-hackathon/dod-agent-traces
收藏资源简介:
该数据集存储了由“DOD - Deploy or Draw”生成的轻量级代理追踪数据,这是一个为Build Small Hackathon构建的多人UNO风格游戏。追踪数据旨在展示游戏AI部分在实时游戏中的参与情况:- **Nemotron机器人回合**:AI对手接收当前棋盘状态,并选择出牌或抽牌。- **IT Director反应**:LLM根据刚出的牌和当前生产危机生成简短的上下文反应。这些记录不是用于微调模型的训练数据,而是运行时追踪数据,记录了AI代理在游戏循环中的使用方式。数据集包含JSONL格式的追踪记录,每个记录包括唯一ID、时间戳、事件类型、模型信息、状态、输入输出数据以及元数据。事件类型主要分为nemotron_turn(记录Nemotron的游戏决策)和director_reaction(记录IT Director的反应生成),并详细说明了各种可能的状态(如成功应用、回退处理或错误情况)。数据集注重隐私保护,仅包含游戏相关数据,不存储个人敏感信息。
This dataset stores lightweight agent tracing data generated by "DOD - Deploy or Draw", a multiplayer UNO-style game built for the Build Small Hackathon. The tracing data aims to demonstrate the participation of game AI components in real-time gameplay: - **Nemotron Bot Turns**: The AI opponent receives the current board state and selects a card to play or draw. - **IT Director Reactions**: The LLM generates brief contextual reactions based on the newly played card and the ongoing production crisis. These records are not training data for model fine-tuning, but runtime tracing data that documents how AI agents are utilized within the game loop. The dataset contains tracing records in JSONL format, with each record including a unique ID, timestamp, event type, model information, state, input-output data, and metadata. The main event types are `nemotron_turn` (recording Nemotron's in-game decision-making) and `director_reaction` (recording the IT Director's reaction generation), with detailed explanations of various possible states such as successful application, fallback processing, or error scenarios. The dataset prioritizes privacy protection, only containing game-related data and not storing any personal sensitive information.




