internlm/RNGBench-Game-Trajectories
收藏资源简介:
RNGBench-Game-Trajectories是一个用于监督微调(SFT)的轨迹数据集,作为RNGBench(重构性非马尔可夫游戏)评估框架的组成部分。该框架旨在测试多模态语言模型是否能够从记忆中重构隐藏状态,并在封闭循环环境中基于这些状态采取行动(即记住并行动场景,其中当前观察不足以决策,模型必须回忆相关历史)。数据集涵盖两个游戏:Matching Pairs(静态分类隐藏状态游戏,如记忆卡匹配游戏)和3D Maze(动态空间隐藏状态游戏,如第一人称视角迷宫导航)。对于每个游戏,提供两种多轮代理轨迹:Optimal(基于规则的最优策略生成)和Roll(从强模型如Kimi-K2.5和Qwen3.5-397B收集的滚动轨迹)。数据以LLaMA-Factory的ShareGPT格式存储,包括图像和文本模态,部分为纯文本。
RNGBench-Game-Trajectories is a supervised fine-tuning (SFT) trajectory dataset accompanying RNGBench (Reconstructive Non-Markov Games), an evaluation framework that tests whether multimodal language models can reconstruct hidden state from memory and act on it in closed-loop environments (the remember-to-act setting, where the current observation alone is insufficient and the model must recall relevant history before deciding). The dataset covers two games: Matching Pairs (a static, categorical hidden-state game, such as a memory card-matching game) and 3D Maze (a dynamic, spatial hidden-state game, like egocentric first-person maze navigation). For each game, it includes two types of multi-turn agent trajectories: Optimal (produced by a rule-based optimal policy) and Roll (rollout trajectories collected from strong models like Kimi-K2.5 and Qwen3.5-397B). The data is stored in LLaMA-Factorys ShareGPT format, with multimodal (image+text) and text-only subsets.




