LudicIndex/ludic-index-turn-logs

该数据集名为Ludic Index — Released Turn Logs，与论文Ludic Index: A Decomposition Framework for Value Loss in Constrained LLM Agents相关。数据集包含经验记录，如回合日志、法官重新评分数据和每个实验的汇总数据。涵盖了Kuhn扑克、Frozen Lake 6x6和RockSample 4x4 POMDP的实验，包括详细的每回合记录、汇总数据和执行元数据。数据集还包括法官调用和比较，总共有6,440个回合和3,957次法官调用。模型面板由通过OpenAI Chat Completions API访问的各种GPT模型组成。数据集采用CC-BY-4.0许可，旨在重现论文的结果。

This dataset is named Ludic Index — Released Turn Logs and is associated with the paper Ludic Index: A Decomposition Framework for Value Loss in Constrained LLM Agents. The dataset includes empirical records such as turn logs, judge rescoring data, and per-experiment aggregates. It covers experiments in Kuhn poker, Frozen Lake 6x6, and RockSample 4x4 POMDP, with detailed per-turn records, aggregates, and manifests. The dataset also includes judge calls and comparisons, with a total volume of 6,440 turns and 3,957 judge calls. The model panel consists of various GPT models accessed via the OpenAI Chat Completions API. The dataset is licensed under CC-BY-4.0 and is intended for reproduction of the papers results.