Serbian-PD

# 🇷🇸 Serbian Public Domain 🇷🇸

**Serbian-Public Domain** or **Serbian-PD** is a large collection aiming to aggregate all Serbian monographies and periodicals in the public domain. As of March 2024, it is the biggest Serbian open corpus.

## Dataset summary
The collection contains 1,405 titles making up 156,712,807 words recovered from multiple sources, including Internet Archive and various European national libraries and cultural heritage institutions. Each parquet file has the full text of 2,000 books selected at random.

## Curation method
The composition of the dataset adheres to the criteria for public domain works in the EU and, consequently, all Berne-countries for EU authors: any publication whose author is dead for more than 70 years. Additionally, the initial consolidation of public domain status for cultural heritage operates in the EU under the 2019 Copyright Directive (art. 14).

As of March 2024, to limit rights verification, we have retained exclusively titles published prior to 1884.

The corpus will be expanded at a later stage to encompass late 19th century and early 20th century publications, after checking for public domain validity.

## Uses
The collection aims to expand the availability of open works for the training of Large Language Models. The text can be used for model training and republished without restriction for reproducibility purposes.

The rationales for creation of this collection are multifold:

* **Scientific**: We observe that the closure of training corpora represents a major barrier to AI research. Large language models face a real crisis of reproducibility.
* **Legal**: With the adoption of the AI Act with its obligations in terms of copyright law compliance for the pretraining corpora, the European AI ecosystem will have to change its provenance practices.
* **Cultural**: The linguistic diversity of the European Union is currently underrepresented. Unlike web archives, open, heritage, administrative, or scientific texts are often of high quality: they are long, multilingual, and editorialized publications.
* **Economical**: Today, value capture is concentrated on players whose financial resources are already considerable, allowing them to collect or purchase data at a high price. Making a royalty-free corpus available to as many people as possible frees innovation in uses and minimizes economic dependencies on dominant actors.

## License
The entire collection is in the public domain in all regions. This means that the patrimonial rights of each individual or collective right holders have expired.

There has been a debate for years in Europe over the definition of public domain and the possibility to restrict its use. Since 2019, the EU Copyright Directive states that "Member States shall provide that, when the term of protection of a work of visual art has expired, any material resulting from an act of reproduction of that work is not subject to copyright or related rights, unless the material resulting from that act of reproduction is original in the sense that it is the author's own intellectual creation." (art. 14)

## Future work
This dataset is not a one-time work but will continue to evolve significantly in three directions:
* Expansion of the dataset to the late 19th and early 20th century works and its further enhancement with currently unexploited collections coming from European patrimonial data repositories.
* Correction of computer generated errors in the text. All the texts have been transcribed automatically through the use of Optical Character Recognition (OCR) software. The original files have been digitized over a long time period (since the mid-2000s) and some documents should be. Future versions will strive either to re-OCRize the original text or use experimental LLM models for partial OCR correction.
* Enhancement of the structure/editorial presentation of the original text. Some parts of the original documents are likely unwanted for large scale analysis or model training (header, page count…). Additionally, some advanced document structures like tables or multi-column layout are unlikely to be well-formatted.

## Acknowledgements
The corpus was stored and processed with the generous support of Scaleway. It was built up with the support and concerted efforts of the state start-up LANGU:IA (start-up d’Etat), supported by the French Ministry of Culture and DINUM, as part of the prefiguration of the service offering of the Alliance for Language technologies EDIC (ALT-EDIC).

Corpus collection has been largely facilitated thanks to the open science LLM community insights, cooperation and support (Occiglot, Eleuther AI, OpenLLM France, Allen AI).

# 🇷🇸 塞尔维亚公共领域数据集 🇷🇸

**塞尔维亚公共领域数据集（Serbian-Public Domain，简称Serbian-PD）**是一个旨在收录所有处于公共领域的塞尔维亚专著与期刊的大型数据集。截至2024年3月，它已是规模最大的塞尔维亚开放语料库。

## 数据集概况
该数据集共收录1405部作品，总字数达156,712,807词，数据来源包括互联网档案馆（Internet Archive）以及多家欧洲国家图书馆与文化遗产机构。每个Parquet文件均包含随机选取的2000部图书的完整文本。

## 数据遴选规则
本数据集的内容遴选遵循欧盟公共领域作品的判定标准，同时适用于所有伯尔尼公约成员国中欧盟作者的作品：即作者去世超过70年的已出版作品。此外，欧盟范围内文化遗产作品的公共领域身份认定，需符合2019年《欧盟版权指令（EU Copyright Directive）》第14条的相关规定。

截至2024年3月，为简化版权权属核验流程，我们仅保留1884年之前出版的作品。后续阶段中，在完成公共领域合法性核验后，本语料库将拓展收录19世纪末至20世纪初的出版物。

## 应用场景
本数据集旨在提升开放作品的可获取性，以供大语言模型（Large Language Model，LLM）训练使用。文本可用于模型训练，且为保障实验可重复性，可无限制地重新发布。

构建本数据集的初衷兼具多重维度：
* **学术层面**：当前训练语料库的封闭化已成为人工智能研究的主要障碍之一，大语言模型正面临严峻的可复现性危机。
* **法律层面**：随着《人工智能法案（AI Act）》的通过，预训练语料库需符合版权合规要求，欧洲人工智能生态系统将不得不调整其数据来源获取方式。
* **文化层面**：当前欧盟的语言多样性在语料库中代表性不足。与网络档案不同，开放的文化遗产、行政或学术文本往往具备更高质量——它们篇幅较长、涵盖多语言且经过专业编辑加工。
* **经济层面**：当前数据价值的获取高度集中于财力雄厚的头部企业，这些企业能够以高价收集或采购数据。向尽可能多的群体提供免版权使用费的语料库，能够激发应用层面的创新，并降低对主导企业的经济依赖。

## 版权声明
本数据集的全部内容在全球范围内均属于公共领域，即所有个人或集体版权持有者的财产性权利均已过期。

多年来，欧洲学界与业界围绕公共领域的定义以及限制其使用的可能性一直存在争议。自2019年起，欧盟《版权指令（EU Copyright Directive）》明确规定："各成员国应规定，当视觉艺术作品的保护期届满后，对该作品进行复制所产生的任何材料不受版权或相关权利约束，除非该复制材料具备作者原创性智力成果的属性。"（第14条）

## 未来规划
本数据集并非一次性项目，未来将从三个方向持续迭代优化：
* **语料扩容**：将数据集覆盖范围拓展至19世纪末至20世纪初的作品，并补充来自欧洲文化遗产数据仓库中尚未被利用的馆藏资源。
* **文本纠错**：修正文本中的机器识别错误。所有文本均通过光学字符识别（Optical Character Recognition，OCR）软件自动转录而来，原始文件自2000年代中期起历经多年数字化处理，部分文档存在识别偏差。未来版本将通过重新执行OCR流程，或借助实验性大语言模型对OCR识别结果进行局部校正。
* **结构优化**：完善原始文本的结构化与编辑呈现形式。原始文档中的部分内容（如页眉、页码等）可能不适合大规模分析或模型训练，此外，表格、多栏布局等复杂文档结构往往格式规整度不足。

## 致谢
本语料库的存储与处理工作得到了Scaleway的慷慨支持。数据集的构建依托于法国国家初创企业LANGU:IA（State Startup）的协作与支持，该项目由法国文化部与DINUM资助，作为语言技术联盟EDIC（ALT-EDIC）服务预研的一部分。

语料库的收集工作在很大程度上得益于开放科学大语言模型社区的经验分享、协作与支持（包括Occiglot、Eleuther AI、OpenLLM France、Allen AI等团队）。