CARTE
收藏资源简介:
CARTE是由巴黎综合理工学院等机构创建的文化锚定区域评估基准,旨在评估大语言模型对法国地理知识与区域差异的细粒度推理能力。该数据集包含2,431道多项选择题,覆盖法国13个本土行政区的14个主题领域,如文化、语言、经济和环境,数据来源于人工筛选的法国开放获取机构文档。其构建过程采用基于文档的增强流程,通过Gemini 3 Flash模型生成问题,并经过自动质量过滤确保内容多样性与地域相关性。该数据集主要应用于大语言模型的文化对齐评估,旨在揭示模型在法语区域文化知识上的覆盖差距,并促进对区域特异性语境的理解与推理能力的提升。
CARTE is a culturally anchored regional evaluation benchmark created by institutions including École Polytechnique in Paris and other relevant academic bodies, which aims to assess the fine-grained reasoning abilities of large language models (LLMs) regarding French geographic knowledge and regional disparities. This dataset consists of 2,431 multiple-choice questions, covering 14 thematic domains across 13 metropolitan administrative regions of France, such as culture, language, economy and environment, with its data sourced from manually curated French open-access institutional documents. Its construction adopts a document-augmented workflow: questions are generated via the Gemini 3 Flash model, followed by automatic quality filtering to guarantee content diversity and regional relevance. This benchmark is primarily utilized for cultural alignment evaluation of LLMs, with the goals of revealing the coverage gaps of models in French regional cultural knowledge, and advancing the improvement of contextual understanding and reasoning capabilities for regional specificity.
根据您提供的数据集详情页面地址(https://huggingface.co/datasets/ScarAlcar/CARTE)及其README文件内容,此数据集的基本信息如下:
- 数据集名称:CARTE
- 许可证:CC BY-NC 4.0(知识共享-署名-非商业使用 4.0 国际许可协议)

- 1CARTE: A Benchmark for Mapping Language Model Knowledge Across France巴黎综合理工学院; 雅典国立技术大学; 穆罕默德·本·扎耶德人工智能大学 · 2026年



