PLUGH
收藏资源简介:
PLUGH数据集由Inworld.AI创建,旨在评估大型语言模型在空间理解和推理方面的能力。该数据集包含从48个不同游戏中提取的125个文本片段,代表了61个不同的空间图。数据集的创建过程涉及使用Jericho引擎重放游戏攻略,以获取游戏文本和空间图。PLUGH数据集主要应用于人工智能领域,特别是用于测试和改进模型在处理空间结构描述文本时的表现。
The PLUGH Dataset, developed by Inworld.AI, is designed to assess the spatial comprehension and reasoning abilities of large language models (LLMs). It comprises 125 text segments extracted from 48 distinct games, which collectively represent 61 unique spatial graphs. The dataset was constructed by replaying game walkthroughs via the Jericho engine to acquire game text and spatial graph data. This dataset is primarily utilized in the field of artificial intelligence, particularly for testing and improving the performance of models when processing text describing spatial structures.
PLUGH 数据集概述
摘要
PLUGH 是一个现代基准测试,包含 5 个任务,每个任务有 125 个输入文本,这些文本从 48 个不同的游戏中提取,代表了 61 个不同的(非同构)空间图,用于评估大型语言模型(LLMs)的空间理解和推理能力。评估显示,尽管某些商业 LLMs 表现出强大的推理能力,但开源竞争对手也能展示出几乎相同的质量水平;然而,所有模型仍有显著的改进空间。
内容
plugh.json:基准数据。sample_responses.py:用于通过 OpenAI API 查询模型或生成提示,以便用于查询任何自定义模型。plugh.responses.json:已采样的多个模型的响应。calc_metrics.py:用于处理响应、解析它们并计算基准指标的脚本。
结果
运行 python calc_metrics.py -i plugh.json -r plugh.responses.json 应生成以下报告:
task1
task1_strict_nodes_f1
| 模型 | 项目数 | 0-shot | 1-shot | 2-shot | 3-shot |
|---|---|---|---|---|---|
| openai:gpt-3-5-turbo | 113 | 55.2% | 66.6% | 70.7% | 73.5% |
| openai:gpt-4-turbo-2024-04-09 | 113 | 73.2% | 76.5% | 76.4% | 77.1% |
| anthropic:claude-3-opus-20240229 | 113 | 70.0% | 77.5% | 78.9% | 81.4% |
| local:llama3_8b | 113 | 64.8% | 60.1% | 65.4% | 67.1% |
| local:mixtral_8x7b | 113 | 25.2% | 67.2% | 69.0% | 69.0% |
| local:llama3_70b | 113 | 68.0% | 73.7% | 74.0% | 74.2% |
| local:mixtral_8x22b | 113 | 67.7% | 75.3% | 76.6% | 75.4% |
| openai:gpt-4o-2024-05-13 | 113 | 71.7% | 73.4% | 74.7% | 75.8% |
task1_strict_edges_f1
| 模型 | 项目数 | 0-shot | 1-shot | 2-shot | 3-shot |
|---|---|---|---|---|---|
| openai:gpt-3-5-turbo | 113 | 30.6% | 42.0% | 46.0% | 49.5% |
| openai:gpt-4-turbo-2024-04-09 | 113 | 53.1% | 56.4% | 56.4% | 57.4% |
| anthropic:claude-3-opus-20240229 | 113 | 50.7% | 59.0% | 60.4% | 62.9% |
| local:llama3_8b | 113 | 41.1% | 36.8% | 40.3% | 40.9% |
| local:mixtral_8x7b | 113 | 12.6% | 41.4% | 42.4% | 44.6% |
| local:llama3_70b | 113 | 46.5% | 53.1% | 53.6% | 54.2% |
| local:mixtral_8x22b | 113 | 45.8% | 54.1% | 56.2% | 55.4% |
| openai:gpt-4o-2024-05-13 | 113 | 52.2% | 53.7% | 56.1% | 57.4% |
task1_fuzzy_nodes_f1
| 模型 | 项目数 | 0-shot | 1-shot | 2-shot | 3-shot |
|---|---|---|---|---|---|
| openai:gpt-3-5-turbo | 113 | 64.6% | 75.8% | 81.3% | 82.8% |
| openai:gpt-4-turbo-2024-04-09 | 113 | 82.9% | 86.4% | 86.3% | 86.8% |
| anthropic:claude-3-opus-20240229 | 113 | 80.3% | 87.1% | 88.3% | 90.7% |
| local:llama3_8b | 113 | 77.5% | 70.3% | 76.2% | 78.6% |
| local:mixtral_8x7b | 113 | 29.2% | 76.7% | 79.1% | 79.3% |
| local:llama3_70b | 113 | 78.9% | 84.0% | 85.4% | 84.9% |
| local:mixtral_8x22b | 113 | 78.9% | 85.2% | 86.5% | 84.7% |
| openai:gpt-4o-2024-05-13 | 113 | 82.0% | 83.3% | 84.3% | 85.5% |
task1_fuzzy_edges_f1
| 模型 | 项目数 | 0-shot | 1-shot | 2-shot | 3-shot |
|---|---|---|---|---|---|
| openai:gpt-3-5-turbo | 113 | 41.4% | 54.3% | 60.4% | 62.3% |
| openai:gpt-4-turbo-2024-04-09 | 113 | 67.1% | 71.1% | 71.3% | 71.8% |
| anthropic:claude-3-opus-20240229 | 113 | 64.4% | 74.3% | 76.2% | 79.0% |
| local:llama3_8b | 113 | 55.7% | 50.9% | 53.9% | 57.2% |
| local:mixtral_8x7b | 113 | 17.0% | 53.0% | 55.6% | 58.4% |
| local:llama3_70b | 113 | 61.6% | 68.0% | 70.1% | 70.5% |
| local:mixtral_8x22b | 113 | 60.1% | 68.5% | 70.6% | 68.8% |
| openai:gpt-4o-2024-05-13 | 113 | 66.8% | 67.8% | 70.8% | 72.3% |
task2a
task2a_strict_distance
| 模型 | 项目数 | 0-shot | 1-shot | 2-shot | 3-shot |
|---|---|---|---|---|---|
| openai:gpt-3-5-turbo | 113 | 38.0% | 31.6% | 29.7% | 28.3% |
| openai:gpt-4-turbo-2024-04-09 | 113 | 26.1% | 12.8% | 12.2% | 11.7% |
| anthropic:claude-3-opus-20240229 | 113 | 34.6% | 11.2% | 10.5% | 10.0% |
| local:llama3_8b | 113 | 80.6% | 65.8% | 53.7% | 54.0% |
| local:mixtral_8x7b | 113 | 61.4% | 46.6% | 47.7% | 46.3% |
| local:llama3_70b | 113 | 58.7% | 15.1% | 14.5% | 15.9% |
| local:mixtral_8x22b | 113 | 34.7% | 19.7% | 16.9% | 18.1% |
| openai:gpt-4o-2024-05-13 | 113 | 21.5% | 10.5% | 10.0% | 9.5% |
task2a_fuzzy_distance
| 模型 | 项目数 | 0-shot | 1-shot | 2-shot | 3-shot |
|---|---|---|---|---|---|
| openai:gpt-3-5-turbo | 113 | 33.6% | 30.5% | 28.5% | 28.2% |
| openai:gpt-4-turbo-2024-04-09 | 113 | 16.2% | 12.8% | 11.9% | 11.7% |
| anthropic:claude-3-opus-20240229 | 113 | 19.2% | 11.1% | 10.5% | 10.0% |
| local:llama3_8b | 113 | 38.7% | 51.3% | 44.1% | 46.4% |
| local:mixtral_8x7b | 113 | 35.9% | 36.6% | 38.2% | 37.6% |
| local:llama3_70b | 113 | 23.4% | 14.9% | 14.4% | 15.8% |
| local:mixtral_8x22b | 113 | 18.7% | 17.7% | 16.4% | 17.6% |
| openai:gpt-4o-2024-05-13 | 113 | 12.7% | 9.5% | 9.3% | 9.3% |
task2b
task2b_strict_distance
| 模型 | 项目数 | 0-shot | 1-shot | 2-shot | 3-shot |
|---|---|---|---|---|---|
| openai:gpt-3-5-turbo | 113 | 64.5% | 61.0% | 58.8% | 59.6% |
| openai:gpt-4-turbo-2024-04-09 | 113 | 22.8% | 18.2% | 15.2% | 15.5% |
| anthropic:claude-3-opus-20240229 | 113 | 31.1% | 23.6% | 21.1% | 21.0% |
| local:llama3_8b | 113 | 90.7% | 69.6% | 67.1% | 66.1% |
| local:mixtral_8x7b | 113 | 83.7% | 60.7% | 63.2% | 60.6% |
| local:llama3_70b | 113 | 59.9% | 32.2% | 30.3% | 32.9% |
| local:mixtral_8x22b | 113 | 63.0% | 40.9% | 39.8% | 39.0% |
| openai:gpt-4o-2024-05-13 | 113 | 21.8% | 13.0% | 11.7% | 12.4% |
task2b_fuzzy_distance
| 模型 | 项目数 | 0-shot | 1-shot | 2-shot | 3-shot |
|---|---|---|---|---|---|
| openai:gpt-3-5-turbo | 113 | 64.5% | 60.2% | 58.4% | 59.5% |
| openai:gpt-4-turbo-2024-04-09 | 113 | 21.7% | 18.2% | 15.2% | 15.5% |
| anthropic:claude-3-opus-20240229 | 113 | 25.7% | 23.6% | 21.1% | 21.0% |
| local:llama3_8b | 113 | 65.0% | 63.6% | 62.2% | 61.7% |
| local:mixtral_8x7b | 113 | 60.8% | 54.4% |




