EscapeBench
收藏资源简介:
EscapeBench是一个用于测试语言模型创造力的基准数据集。它包括基准数据、BaseAgent和EscapeAgent的实现以及运行测试的脚本。
EscapeBench is a benchmark dataset for evaluating the creativity of language models. It comprises the benchmark data, the implementations of BaseAgent and EscapeAgent, as well as the scripts for running the tests.
EscapeBench 数据集概述
数据集简介
EscapeBench 是一个用于测试语言模型创造力的基准数据集。该数据集包含基准数据、BaseAgent 和 EscapeAgent 的实现,以及运行测试的脚本。
数据集结构
-
src/ 目录:
agent_base.py: BaseAgent 的实现agent_creative.py: EscapeAgent 的实现human.py: 人类玩家接口env/: 核心游戏引擎设计,包括场景、物品、工具等
-
data/ 目录:
<game>.yaml: 不同难度级别的游戏设置文件check_data.py: 检查游戏数据逻辑错误的脚本reference/: 成功完成游戏的动作链,作为提示参考
游戏设置
每个游戏设置的数据逻辑如下: yaml
- name: <scene name>
desc: <scene description>
scene_relations:
<prompt>: <nearby scene name>
...
items:
- position: <position of item>
item:
name: <item name>
interactable: <True/False>
visible: <True/False>
states:
- desc: <item description>
neg_reward: <negative env feedback if wrong action is tried>
transitions:
- wait_for:
- <waited action> [click], [apply <tool name>], [input <str>] trigger:
- <trigger effect> [change_visible, scene/item/tool, <name>, True/False], [change_interact, item, <name>, True/False], [change_state, item/tool, <name>, <int>], [become_tool, <name>] reward: <positive env feedback if correct action is performed> tools:
- wait_for:
- position: <position of tool> tool: name: fragment visible: <True/False> states: - desc: <tool description> [apply_to/wait_for]: - <tool name>
- desc: <item description>
neg_reward: <negative env feedback if wrong action is tried>
transitions:
- position: <position of item>
item:
name: <item name>
interactable: <True/False>
visible: <True/False>
states:
排行榜
当前 EscapeBench 性能排行榜如下:
| 排名 | 代理模型 | 提示使用次数 | 总步数 |
|---|---|---|---|
| 1 | Claude-3.5-Sonnet | 8.97 | 690.31 |
| 2 | GPT-4o | 10.30 | 723.61 |
| 3 | Gemini-1.5-pro | 11.06 | 824.31 |
| 4 | Llama-3.1-70B | 14.53 | 982.42 |
| 5 | GPT-4o-mini | 15.19 | 1002.39 |
| 6 | Qwen2.5-72B | 16.50 | 1102.50 |
| 7 | Yi-1.5-34B | 24.00 | 1573.33 |
| 8 | Ministral | 25.31 | 1556.97 |
| 9 | DeepSeek-LLM-67B | 25.50 | 1558.47 |
| 10 | Llama-3.1-8B | 25.86 | 1543.30 |
引用
text @article{qian2024escapebench, title={EscapeBench: Pushing Language Models to Think Outside the Box}, author={Qian, Cheng and Han, Peixuan and Luo, Qinyu and He, Bingxiang and Chen, Xiusi and Zhang, Yuji and Du, Hongyi and Yao, Jiarui and Yang, Xiaocheng and Zhang, Denghui and Li, Yunzhu and Ji, Heng}, journal={arXiv preprint arXiv:2412.13549}, year={2024} }




