arc-agi3-ka59-agent-trajectories
收藏资源简介:
ARC-AGI-3 ka59 代理轨迹数据集包含来自大型语言模型(LLM)代理(使用Claude Code驱动文件中继工具)玩ARC-AGI-3游戏ka59的完整游戏轨迹,作为ARA-as-world-model泛化实验的一部分。该代理在游戏过程中实时构建结构化世界模型(称为Agent-Native Research Artifact,ARA),并利用该模型解决无法通过初始探索破解的关卡。数据集在游戏过程中持续同步更新,每次世界模型固化时都会更新。数据内容涵盖多个方面:代理的世界模型和推理跟踪(位于reasoning/目录);每个子代理的动作-推理跟踪,以每行JSON格式记录动作及其原因(traces/目录);游戏进度快照、机制分析和获胜动作脚本(solutions/目录);会话级的帧记录(recordings/目录);回合级事件日志(episodes/目录);工具日志和自由形式会话笔记(logs/目录);ARC API记分卡(scorecards/目录);世界模型预测账本,包含咨询置信度和后续验证(predictions.jsonl文件);游戏的资源使用指标和ARA增长曲线(accounting/目录);以及消融或展示工件(demo/目录)。此外,数据集包括会话索引(sessions_index.json)和来源清单(manifest.json),后者记录代理模型、仓库提交、协议哈希和同步时间等信息。仓库标签L<n>-cleared标记了每个关卡清除时的完整记录快照,适用于下游代理评估实验。数据集适用于强化学习、游戏玩法分析、LLM代理行为研究、世界模型构建和任务泛化等任务,但未明确指定样本数量或模态细节。
The ARC-AGI-3 ka59 agent trajectory dataset contains complete game trajectories from a large language model (LLM) agent (Claude Code-driven file relay tool) playing the ARC-AGI-3 game ka59, as part of the ARA-as-world-model generalization experiment. The agent constructs a structured world model (called Agent-Native Research Artifact, ARA) in real-time during gameplay and uses it to solve levels that cannot be cracked through initial exploration. The dataset is continuously synchronized during gameplay, updated each time the world model solidifies. The data covers multiple aspects: the agents world model and reasoning traces (in the reasoning/ directory); action-reasoning traces for each sub-agent, recorded in per-line JSON format with actions and their reasons (traces/ directory); game progress snapshots, mechanism analysis, and winning action scripts (solutions/ directory); session-level frame recordings (recordings/ directory); turn-level event logs (episodes/ directory); tool logs and free-form session notes (logs/ directory); ARC API scorecards (scorecards/ directory); world model prediction ledgers with consultation confidence and subsequent validation (predictions.jsonl file); resource usage metrics for the game and ARA growth curves (accounting/ directory); and ablation or demonstration artifacts (demo/ directory). Additionally, the dataset includes a session index (sessions_index.json) and a provenance manifest (manifest.json), which records information such as the agent model, repository commits, protocol hashes, and synchronization times. Repository tags L<n>-cleared mark complete recording snapshots at each level clearance, suitable for downstream agent evaluation experiments. The dataset is applicable for tasks like reinforcement learning, gameplay analysis, LLM agent behavior research, world model construction, and task generalization, but does not explicitly specify sample counts or modality details.
数据集概述
数据集名称:ARC-AGI-3 ka59 — Agent Trajectories
许可证:CC-BY-4.0
语言:英语
任务类别:强化学习、其他
标签:ARC-AGI, ARC-AGI-3, 智能体轨迹, 游戏玩法, 大语言模型智能体, 世界模型, Claude
数据集描述:
该数据集包含一个LLM智能体(由Claude Code驱动的文件中继框架)在玩ARC-AGI-3游戏ka59时的游戏轨迹。这些轨迹属于ARA作为世界模型的泛化实验的一部分。智能体在游戏过程中实时构建结构化的世界模型(Agent-Native Research Artifact),并在无法通过冷探索解决的关卡中参考该模型。配套数据集为:arc-agi3-ls20-agent-trajectories。数据在游戏过程中持续同步——每次世界模型的具体化都会触发推送。
数据集结构:
reasoning/ara-ka59/:智能体的世界模型及推理轨迹(包含logic/、trace/、staging/子目录)traces/:每个子智能体的动作-推理轨迹,每行一个JSON动作及原因说明solutions/GAME.md:状态快照,包含进度、机制、陷阱、暂停原因solutions/replay/:每个关卡的获胜动作脚本(L<n>.txt)recordings/:按帧的记录,每个会话一个文件(<game>-<hash>.<agent>.<guid>.recording.jsonl)episodes/:回合级的情节日志(跨运行拼接)logs/:框架日志及自由形式的会话笔记(git历史保留之前的会话)scorecards/:ARC API评分卡(保存时)predictions.jsonl:世界模型预测记录,包含置信度及后续判定accounting/:该游戏的分类账行(令牌/步骤)、指标部分、ARA增长曲线demo/:消融/展示工件(当游戏有时)sessions_index.json:每个记录会话的条目(智能体、GUID、帧数、最终状态)manifest.json:来源信息,包括智能体模型、仓库提交、协议哈希、同步时间
其他信息:
仓库标签 L<n>-cleared 标记每个关卡通关时的完整记录,可作为下游智能体评估实验的持出快照。智能体决策时间(动作间间隔)、死亡分析及意外分析可从记录和轨迹中推导,未单独存储。




