GRUFF
收藏资源简介:
GRUFF是由汉堡大学可信人工智能实验室等机构创建的首个大规模德语代词忠实度评估数据集,旨在研究语言模型在德语中正确跟踪和复现代词的能力。该数据集覆盖四种名词一致性系统和四组代词,包含超过七百万个独特实例,通过模板组合生成,数据来源于职业-参与者对的德语翻译及语法结构模板。其创建过程涉及将英语RUFF模板翻译并重构以适应德语语法复杂性,包括名词系统与代词集的配对及人工验证。该数据集应用于自然语言处理领域,专门用于评估语言模型在德语中的代词忠实度、推理能力及性别偏见,旨在推动性别包容性语言和指代推理研究,解决模型在复杂语境下代词使用的准确性与公平性问题。
GRUFF is the first large-scale German pronoun faithfulness evaluation dataset developed by the Trusted Artificial Intelligence Laboratory at the University of Hamburg and other institutions. It is designed to investigate the ability of language models to correctly track and resolve pronouns in German. The dataset covers four noun agreement systems and four sets of pronouns, contains over seven million unique instances, and is generated via template combinations, with its data sourced from German translations of career-participant pairs and grammatical structure templates. Its creation process involved translating and reconstructing English RUFF templates to accommodate the grammatical complexity of German, including the pairing of noun systems and pronoun sets, as well as manual validation. This dataset is applied in the field of natural language processing, specifically for evaluating the pronoun faithfulness, reasoning capabilities and gender bias of language models in German. It aims to advance research on gender-inclusive language and referential reasoning, and address the issues of accuracy and fairness in pronoun usage by models in complex contextual environments.
数据集概述
数据集名称:GRUFF
地址:https://github.com/TAI-HAMBURG/gruff
语言:德语
类型:德语代词评估数据集
主要用途
- 生成德语代词评估项目
- 添加显性和隐性上下文干扰项
- 抽取平衡的评估子集
- 对语言模型在二元和非二元代词形式上进行评分
仓库内容
data/:任务和上下文生成的源TSV文件task.tsvcontext.tsv
scripts/:用于生成、采样和模型评分的Python脚本- 包括
baseline_task.py、add_context.py、sample_templates.py、sample_for_humans.py、score_models.py等
- 包括
核心工作流程
从项目根目录执行以下脚本管线:
python scripts/baseline_task.pypython scripts/add_context.py tasks_no_context.tsv data/context.tsvpython scripts/sample_templates.pypython scripts/sample_for_humans.pypython scripts/score_models.py 11_*.tsv 13_*.tsv 15_*.tsv
重要注意事项
在分享或发布仓库前,确保scripts/constants.py中不包含真实的Hugging Face访问令牌。
其他文档
脚本的详细描述见scripts/README.md。




