OpenLID-v3
收藏资源简介:
OpenLID-v3是OpenLID-v2数据集的更新版本,专为训练高覆盖率的语言识别模型而设计。该数据集覆盖194种语言变体及一个非语言类别,其语言变体标签与FLORES+评估基准兼容。每个数据实例包含三个字段:`text`(文本行)、`language`(由ISO 639-3语言代码和ISO 15924文字代码组成的语言标签,例如spa_Latn)、`source`(表示数据来源的标签)。数据集仅提供训练分割。该资源旨在解决现有语言识别工具在区分密切相关的语言以及过滤噪声方面的挑战,特别关注了三组密切相关的语言(如波斯尼亚语、克罗地亚语和塞尔维亚语;意大利北部和法国南部的罗曼语变体;斯堪的纳维亚语言),并贡献了新的评估数据集。使用该数据集时需注意:它涵盖了许多服务不足的语言,但由于数据量和领域的限制,应避免过度宣称其性能或覆盖范围;语言识别本身是一种规范性活动,可能边缘化少数民族方言、文字或微语言,且模型在不同语言上的性能不均衡,可能影响下游任务。数据集许可允许非商业用途的开放使用,具体许可信息在仓库的`licenses`目录中。
OpenLID-v3 is an updated version of the OpenLID-v2 dataset, designed for training high-coverage language identification models. It covers 194 language varieties and one non-language category, with language variety labels compatible with the FLORES+ evaluation benchmark. Each data instance includes three fields: `text` (text line), `language` (a language label composed of ISO 639-3 language code and ISO 15924 script code, e.g., spa_Latn), and `source` (a label indicating the data source). The dataset provides only a training split. This resource aims to address challenges in existing language identification tools, such as distinguishing closely related languages and filtering noise, with a particular focus on three groups of closely related languages (e.g., Bosnian, Croatian, and Serbian; Romance varieties in Northern Italy and Southern France; Scandinavian languages), and contributes new evaluation datasets. When using this dataset, note that it includes many underserved languages, but due to limitations in data volume and domain, overclaiming its performance or coverage should be avoided; language identification itself is a normative activity that may marginalize minority dialects, scripts, or microlanguages, and model performance varies across languages, which may affect downstream tasks. The dataset license permits open use for non-commercial purposes, with specific license information in the repositorys `licenses` directory.
数据集概述
OpenLID-v3 是一个用于语言识别(Language Identification, LID)的数据集,是 OpenLID-v2 的更新版本。
基本信息
- 数据集名称: OpenLID-v3
- 许可证: 其他 (详见 licenses 目录)
- 任务: 文本分类 (text-classification)
- 语言覆盖: 194 种语言变体 + 非语言类别(not-a-language class),语言标签兼容 FLORES+ 评估基准
- 仓库: https://github.com/hplt-project/openlid
- 论文: OpenLID-v3: Improving the Precision of Closely Related Language Identification – An Experience Report
数据集结构
-
数据实例: 每条数据包含三个字段:
text: 文本内容language: 由 ISO 639-3 语言代码 + ISO 15924 文字代码组成的语言标签(例如 "spa_Latn")source: 数据来源标签(例如 "lti")
-
数据划分: 仅提供训练集(train split)
使用与引用
-
加载方式: python from datasets import load_dataset ds = load_dataset(HPLT/OpenLID-v3, split=train)
-
训练脚本: 提供 训练脚本 用于准备数据集以训练语言识别模型
-
引用信息: 使用该数据集时需引用相关论文及来源数据集的作者,详见 citations.bib 文件
注意事项
- 社会影响: 数据集覆盖了许多低资源语言,但数据量和领域有限,需注意不要夸大性能或覆盖范围
- 偏见讨论: 语言识别本身具有规范性,可能排除少数方言、文字或微语言;模型在不同语言上的性能不均衡,可能影响特定群体的下游任务表现
- 许可信息: 各来源数据的许可证详见 licenses 目录,所有许可证均覆盖非商业用途的开放使用




