A multilingual annotated dataset for metadata extraction from Persian, Arabic and English scientific articles: header and reference section annotations Authors: Arash Yousefi Jordehi, Reza Fallah, Ebrahim Nikmard, Mahsa Hosseini Khasheh Heyran, Seyed Abolghasem Mirroshandel Published: 2026 ABSTRACT This data article describes a multilingual annotated dataset for training and evaluating metadata extraction systems for scientific articles written in Persian, Arabic and English. The dataset consists of 635 JSON files, one per scientific article, holding the annotations of the article header and of the reference section. The article-level language distribution is 352 Persian (55.4 %), 142 English (22.4 %) and 141 Arabic (22.2 %). Of the 635 articles, 621 contain annotated reference sections; the remaining 14 contain header annotations only. In total, 17,832 individual reference strings are annotated, comprising 8476 Persian (47.5 %), 5741 Arabic (32.2 %) and 3615 English (20.3 %) strings. Header annotations cover seven metadata fields: title, authors, affiliations, abstract, keywords, venue information and DOI. Reference annotations cover twelve metadata fields: title, authors, journal or conference name, volume, issue number, page range, year, DOI, publisher, book identifier, organisation and reference type, alongside the raw reference string. The source articles were retrieved from open-access repositories covering >580 conferences and journals. Annotation was carried out over a period of more than twelve months by a team of ten annotators using a purpose-built Java labelling tool with color-coded tagging, and every annotated file was checked against its source PDF in a dedicated review phase. Each file is self-contained: it stores the annotated text spans verbatim together with their labels, so the original PDF is not required in order to use the data. The format can be consumed directly by sequence labelling methods such as conditional random fields, recurrent neural networks, transformer-based token classifiers and large language models. The data can be reused to develop and benchmark metadata extraction and citation parsing systems and to support indexing in digital libraries and scholarly search services covering Persian and Arabic literature. Specifications TableSubjectComputer ScienceSpecific subject areaNatural language processing; document analysis and recognition; extraction of header and reference metadata from multilingual (Persian, Arabic, English) scientific articles.Type of dataText (JSON files); Table (summary statistics reported in this article); Figure. Analysed, manually annotated by trained human annotators. The raw reference strings are retained verbatim inside every record.Data collectionOpen-access scientific articles in Persian, Arabic and English were downloaded in PDF form from public online repositories covering >580 conferences and journals. Each PDF was checked to confirm that it is a scientific article and was classified by language and source venue. Header and reference metadata were then labelled by ten trained annotators over more than twelve months using a purpose-built Java labelling tool with color-coded tags, and exported as structured JSON. Every exported file was compared against its source PDF in a separate review phase; files with discrepancies were returned for re-annotation. The published dataset comprises 635 annotated files containing 17,832 annotated reference strings. The original PDF files are not redistributed because their copyright is held by the respective publishers.Data source locationInstitution: Department of Computer Engineering, Faculty of Engineering, University of Guilan. City/Town/Region: Rasht, Guilan Province. Country: Iran. Latitude and longitude: 37.2808° N, 49.5832° E. The annotation work was carried out at the above location. The source articles themselves were obtained from publicly accessible online repositories and are not tied to a single geographic location; their identifiers are listed in the repository.Data accessibilityRepository name: Mendeley Data. Data identification number (DOI): 10.17632/bk52kn69bh.1. Direct URL to data: https://data.mendeley.com/datasets/bk52kn69bh/1. Licence: CC BY 4.0. Instructions for accessing these data: the dataset is publicly accessible without registration or prior request. Download the archive, which contains the 635 annotated JSON files, a README file, the annotation schema and a dataset statistics report. Citation: see reference [1].Related research articleNone SubjectComputer ScienceSpecific subject areaNatural language processing; document analysis and recognition; extraction of header and reference metadata from multilingual (Persian, Arabic, English) scientific articles.Type of dataText (JSON files); Table (summary statistics reported in this article); Figure. Analysed, manually annotated by trained human annotators. The raw reference strings are retained verbatim inside every record.Data collectionOpen-access scientific articles in Persian, Arabic and English were downloaded in PDF form from public online repositories covering >580 conferences and journals. Each PDF was checked to confirm that it is a scientific article and was classified by language and source venue. Header and reference metadata were then labelled by ten trained annotators over more than twelve months using a purpose-built Java labelling tool with color-coded tags, and exported as structured JSON. Every exported file was compared against its source PDF in a separate review phase; files with discrepancies were returned for re-annotation. The published dataset comprises 635 annotated files containing 17,832 annotated reference strings. The original PDF files are not redistributed because their copyright is held by the respective publishers.Data source locationInstitution: Department of Computer Engineering, Faculty of Engineering, University of Guilan. City/Town/Region: Rasht, Guilan Province. Country: Iran. Latitude and longitude: 37.2808° N, 49.5832° E. The annotation work was carried out at the above location. The source articles themselves were obtained from publicly accessible online repositories and are not tied to a single geographic location; their identifiers are listed in the repository.Data accessibilityRepository name: Mendeley Data. Data identification number (DOI): 10.17632/bk52kn69bh.1. Direct URL to data: https://data.mendeley.com/datasets/bk52kn69bh/1. Licence: CC BY 4.0. Instructions for accessing these data: the dataset is publicly accessible without registration or prior request. Download the archive, which contains the 635 annotated JSON files, a README file, the annotation schema and a dataset statistics report. Citation: see reference [1].Related research articleNone Value of the Data •The dataset supplies ready-to-use supervised training material for metadata extraction from Persian and Arabic scientific articles. Openly annotated resources for these two languages are scarce, and the tools most widely used in this area, namely GROBID [2], ParsCit [3] and CERMINE [4], were developed and evaluated on English material. Groups working on non-English scholarly document processing can therefore train and evaluate models without first having to build an annotated corpus of their own.•Header and reference annotations are provided for the same set of articles, which allows a complete extraction pipeline, from the identification of bibliographic fields on the first page through to the parsing of individual reference strings, to be trained and evaluated on a single resource. Seven header fields are annotated across 635 articles and twelve reference fields across 17,832 reference strings.•The source articles were drawn from >580 conferences and journals, so the data span a wide range of page layouts, font conventions, citation styles (including APA, MLA, Chicago, Harvard and Vancouver variants) and language-specific typographic conventions. This variation is useful to anyone who needs to measure how well an extraction model generalizes beyond a single publisher template.•Each record is a self-contained JSON object in which the annotated spans are stored verbatim, so the data can be loaded directly by sequence labelling frameworks (conditional random fields, BiLSTM-CRF), by transformer-based token classifiers, or as prompt and evaluation material for large language models, without PDF parsing or further preprocessing.•Beyond model training, the annotations can be reused by developers of digital libraries, citation indexes, bibliometric tools and scholarly search services that need to process Persian and Arabic literature, and by researchers running cross-lingual comparisons of extraction accuracy.•Every file was checked against its source PDF in a separate review phase, and the annotation guidelines and the JSON schema are distributed with the data, so users can inspect the annotation conventions and extend the corpus consistently if they wish. •The dataset supplies ready-to-use supervised training material for metadata extraction from Persian and Arabic scientific articles. Openly annotated resources for these two languages are scarce, and the tools most widely used in this area, namely GROBID [2], ParsCit [3] and CERMINE [4], were developed and evaluated on English material. Groups working on non-English scholarly document processing can therefore train and evaluate models without first having to build an annotated corpus of their own. The dataset supplies ready-to-use supervised training material for metadata extraction from Persian and Arabic scientific articles. Openly annotated resources for these two languages are scarce, and the tools most widely used in this area, namely GROBID [2], ParsCit [3] and CERMINE [4], were developed and evaluated on English material. Groups working on non-English scholarly document processing can therefore train and evaluate models without first having to build an annotated corpus of their own. •Header and reference annotations are provided for the same set of articles, which allows a complete extraction pipeline, from the identification of bibliographic fields on the first page through to the parsing of individual reference strings, to be trained and evaluated on a single resource. Seven header fields are annotated across 635 articles and twelve reference fields across 17,832 reference strings. Header and reference annotations are provided for the same set of articles, which allows a complete extraction pipeline, from the identification of bibliographic fields on the first page through to the parsing of individual reference strings, to be trained and evaluated on a single resource. Seven header fields are annotated across 635 articles and twelve reference fields across 17,832 reference strings. •The source articles were drawn from >580 conferences and journals, so the data span a wide range of page layouts, font conventions, citation styles (including APA, MLA, Chicago, Harvard and Vancouver variants) and language-specific typographic conventions. This variation is useful to anyone who needs to measure how well an extraction model generalizes beyond a single publisher template. The source articles were drawn from >580 conferences and journals, so the data span a wide range of page layouts, font conventions, citation styles (including APA, MLA, Chicago, Harvard and Vancouver variants) and language-specific typographic conventions. This variation is useful to anyone who needs to measure how well an extraction model generalizes beyond a single publisher template. •Each record is a self-contained JSON object in which the annotated spans are stored verbatim, so the data can be loaded directly by sequence labelling frameworks (conditional random fields, BiLSTM-CRF), by transformer-based token classifiers, or as prompt and evaluation material for large language models, without PDF parsing or further preprocessing. Each record is a self-contained JSON object in which the annotated spans are stored verbatim, so the data can be loaded directly by sequence labelling frameworks (conditional random fields, BiLSTM-CRF), by transformer-based token classifiers, or as prompt and evaluation material for large language models, without PDF parsing or further preprocessing. •Beyond model training, the annotations can be reused by developers of digital libraries, citation indexes, bibliometric tools and scholarly search services that need to process Persian and Arabic literature, and by researchers running cross-lingual comparisons of extraction accuracy. Beyond model training, the annotations can be reused by developers of digital libraries, citation indexes, bibliometric tools and scholarly search services that need to process Persian and Arabic literature, and by researchers running cross-lingual comparisons of extraction accuracy. •Every file was checked against its source PDF in a separate review phase, and the annotation guidelines and the JSON schema are distributed with the data, so users can inspect the annotation conventions and extend the corpus consistently if they wish. Every file was checked against its source PDF in a separate review phase, and the annotation guidelines and the JSON schema are distributed with the data, so users can inspect the annotation conventions and extend the corpus consistently if they wish. Background The growth in the volume of scientific publications has increased the demand for automated metadata extraction, which underpins digital libraries, citation indexes and scholarly search services. For English-language articles a number of mature systems are available, among them GROBID [2], ParsCit [3] and CERMINE [4]. Reference string parsing in particular has been approached with hidden Markov models [[5], [6], [7]], conditional random fields [8,9], structural support vector machines [10], sequence alignment [11] and hierarchical knowledge representation [12], and has been studied for PDF as well as HTML sources [13,14]. Surveys and comparative evaluations of the resulting tools are reported in [[15], [16], [17]]. These efforts have concentrated on English and, to a lesser extent, on other Latin-script languages. We are not aware of a publicly available annotated corpus for header or reference metadata extraction from Persian or Arabic scientific articles, which limits the extent to which scholarly output in these languages can be indexed by the same infrastructure that serves English-language literature. Right-to-left script, mixed-script reference strings, the use of the Solar Hijri calendar in Persian publication years and venue-specific formatting conventions all mean that models trained on English material cannot simply be transferred. The present dataset was assembled to provide the annotated material required for that work. This data article describes the annotated data, its structure, its annotation schema and the procedure by which it was produced, so that the data can be reused and extended by other groups. Data description The dataset is deposited on Mendeley Data [1] and is distributed as a single ZIP archive. The contents of the deposited archive are listed in Table 1.Table 1Contents of the deposited archive.Table 1 dummy alt textFile or directoryContent/JSON_Annotations/635 JSON files, one per scientific article. Each file contains the annotations of the article header and, where present, the annotations of the reference section of the same article. File names follow the pattern metadata_paperN.json where N is an integer identifier./README.mdDescription of the dataset structure, the JSON schema, the annotation conventions and usage instructions./annotation_schema.jsonA JSON Schema document formally defining the structure and data types of every annotation file./dataset_statistics.txtSummary statistics of the released dataset, reproducing the counts reported in this article. Table 1Contents of the deposited archive.Table 1 dummy alt textFile or directoryContent/JSON_Annotations/635 JSON files, one per scientific article. Each file contains the annotations of the article header and, where present, the annotations of the reference section of the same article. File names follow the pattern metadata_paperN.json where N is an integer identifier./README.mdDescription of the dataset structure, the JSON schema, the annotation conventions and usage instructions./annotation_schema.jsonA JSON Schema document formally defining the structure and data types of every annotation file./dataset_statistics.txtSummary statistics of the released dataset, reproducing the counts reported in this article. Contents of the deposited archive. The uncompressed archive occupies approximately 10.5 MB. All files are UTF-8 encoded. Language distribution The 635 annotated articles are distributed across the three languages as shown in Table 2, and the 17,832 annotated reference strings as shown in Table 3. The same distributions are shown graphically in Fig. 1.Table 2Distribution of the 635 annotated articles by language.Table 2 dummy alt textLanguageNumber of articlesPercentagePersian35255.4 %English14222.4 %Arabic14122.2 %Total635100 %Table 3Distribution of the 17,832 annotated reference strings by language. Reference strings are counted in the language in which they are written, which is why the distribution differs from that of Table 2: Persian and Arabic articles frequently cite English-language sources and vice versa.Table 3 dummy alt textLanguageNumber of reference stringsPercentagePersian847647.5 %Arabic574132.2 %English361520.3 %Total17,832100 %Fig. 1Composition of the dataset: (a) the 635 annotated articles by language; (b) the 17,832 annotated reference strings by language; (c) the annotated reference strings by reference type.Fig 1 dummy alt text Table 2Distribution of the 635 annotated articles by language.Table 2 dummy alt textLanguageNumber of articlesPercentagePersian35255.4 %English14222.4 %Arabic14122.2 %Total635100 % Distribution of the 635 annotated articles by language. Table 3Distribution of the 17,832 annotated reference strings by language. Reference strings are counted in the language in which they are written, which is why the distribution differs from that of Table 2: Persian and Arabic articles frequently cite English-language sources and vice versa.Table 3 dummy alt textLanguageNumber of reference stringsPercentagePersian847647.5 %Arabic574132.2 %English361520.3 %Total17,832100 % Distribution of the 17,832 annotated reference strings by language. Reference strings are counted in the language in which they are written, which is why the distribution differs from that of Table 2: Persian and Arabic articles frequently cite English-language sources and vice versa. Composition of the dataset: (a) the 635 annotated articles by language; (b) the 17,832 annotated reference strings by language; (c) the annotated reference strings by reference type. Header annotations Seven metadata fields are annotated in the header of every article. A field is absent from a record when it does not occur in the source article. The coverage of each field over the 635 articles is given in Table 4 and shown in Fig. 2(a).Table 4Header metadata fields and their coverage across the 635 annotated articles.Table 4 dummy alt textFieldDescriptionCoveragetitleFull article title99.4 % (631/635)authorsList of author names98.1 % (623/635)abstractFull abstract text93.4 % (593/635)affiliationsAuthor affiliations and supplementary information93.2 % (592/635)venue_informationJournal or conference name, volume, issue, year89.8 % (570/635)keywordsList of author-defined keywords78.3 % (497/635)doiDOI string, where present in the article15.4 % (98/635)Fig. 2Annotation coverage: (a) the seven header fields, as a percentage of the 635 annotated articles; (b) the reference fields, as a percentage of the 17,832 annotated reference strings. The reference_item field is the raw string and is present by construction.Fig 2 dummy alt text Table 4Header metadata fields and their coverage across the 635 annotated articles.Table 4 dummy alt textFieldDescriptionCoveragetitleFull article title99.4 % (631/635)authorsList of author names98.1 % (623/635)abstractFull abstract text93.4 % (593/635)affiliationsAuthor affiliations and supplementary information93.2 % (592/635)venue_informationJournal or conference name, volume, issue, year89.8 % (570/635)keywordsList of author-defined keywords78.3 % (497/635)doiDOI string, where present in the article15.4 % (98/635) Header metadata fields and their coverage across the 635 annotated articles. Annotation coverage: (a) the seven header fields, as a percentage of the 635 annotated articles; (b) the reference fields, as a percentage of the 17,832 annotated reference strings. The reference_item field is the raw string and is present by construction. Reference annotations Of the 635 articles, 621 contain an annotated reference section; the remaining 14 contain header annotations only, because no reference section could be identified in the source PDF. Each reference string is stored verbatim in the reference_item field and is annotated with up to twelve metadata fields, listed in Table 5. The reference_item field is the raw string rather than an extracted field, which is why its coverage is 100 % by construction and why it is not counted among the twelve annotated fields. Field coverage is shown graphically in Fig. 2(b).Table 5Reference metadata fields and their coverage across the 17,832 annotated reference strings.Table 5 dummy alt textFieldDescriptionCoveragereference_itemFull raw reference string (not an annotated field)100 % (17,832)reference_typeJournal, Conference, Book or Thesis99.6 % (17,767)reference_titleTitle of the cited work97.5 % (17,395)reference_authorsAuthors of the cited work96.8 % (17,256)reference_yearPublication year95.8 % (17,080)reference_journalJournal or conference name49.0 % (8744)reference_pagesPage range46.8 % (8346)reference_volumeVolume number41.2 % (7350)reference_publisherPublisher name34.8 % (6210)reference_numberIssue number31.0 % (5524)reference_organizationOrganisation or institution12.5 % (2234)reference_doiDOI of the cited work12.1 % (2149)reference_book_idBook identifier, such as an ISBN2.3 % (418) Table 5Reference metadata fields and their coverage across the 17,832 annotated reference strings.Table 5 dummy alt textFieldDescriptionCoveragereference_itemFull raw reference string (not an annotated field)100 % (17,832)reference_typeJournal, Conference, Book or Thesis99.6 % (17,767)reference_titleTitle of the cited work97.5 % (17,395)reference_authorsAuthors of the cited work96.8 % (17,256)reference_yearPublication year95.8 % (17,080)reference_journalJournal or conference name49.0 % (8744)reference_pagesPage range46.8 % (8346)reference_volumeVolume number41.2 % (7350)reference_publisherPublisher name34.8 % (6210)reference_numberIssue number31.0 % (5524)reference_organizationOrganisation or institution12.5 % (2234)reference_doiDOI of the cited work12.1 % (2149)reference_book_idBook identifier, such as an ISBN2.3 % (418) Reference metadata fields and their coverage across the 17,832 annotated reference strings. The distribution of reference strings across the four reference types is given in Table 6 and shown in Fig. 1(c). The 65 strings labelled Unknown are those for which the annotators could not determine the type from the string alone.Table 6Distribution of the annotated reference strings by reference type.Table 6 dummy alt textReference typeCountPercentageJournal10,28557.7 %Book581832.6 %Thesis10585.9 %Conference6063.4 %Unknown650.4 %Total17,832100 % Table 6Distribution of the annotated reference strings by reference type.Table 6 dummy alt textReference typeCountPercentageJournal10,28557.7 %Book581832.6 %Thesis10585.9 %Conference6063.4 %Unknown650.4 %Total17,832100 % Distribution of the annotated reference strings by reference type. Record format Each record is a single JSON object. The header fields are stored at the top level and the annotated reference strings are stored in the references array. Persian and Arabic text is stored as UTF-8 in logical order, without additional bidirectional control characters. Fields that do not occur in the source article are stored as null. The structure is illustrated below; the values shown are placeholders.{"pdfType": "PAPER","language": "FARSI","title": "
","abstract": "","keyword": ", ","venue_information": "","authors": ["", ""],"affiliations": ["", ""],"doi": null,"references": [ { "reference_item": "", "reference_title": "", "reference_authors": ["<author name>"], "reference_journal": "<journal name>", "reference_pages": "45-60", "reference_number": "3", "reference_volume": "12", "reference_year": "1398", "reference_publisher": null, "reference_organazation": null, "reference_doi": null, "reference_book_id": null, "reference_type": "Journal" } ]} {"pdfType": "PAPER","language": "FARSI","title": "<article title>","abstract": "<abstract text>","keyword": "<keyword 1>, <keyword 2>","venue_information": "<journal name, year, issue>","authors": ["<first author>", "<second author>"],"affiliations": ["<university ...>", "<email: ...>"],"doi": null,"references": [ { "reference_item": "<full raw reference string>", "reference_title": "<title of the cited work>", "reference_authors": ["<author name>"], "reference_journal": "<journal name>", "reference_pages": "45-60", "reference_number": "3", "reference_volume": "12", "reference_year": "1398", "reference_publisher": null, "reference_organazation": null, "reference_doi": null, "reference_book_id": null, "reference_type": "Journal" } ]} Listing 1. Structure of a single annotated record. Note that the key for the organization field is spelled reference_organazation in the released files; it is documented here exactly as it appears in the data so that parsers written against this article will work without modification. Persian publication years are recorded as they appear in the source article, which means that Persian references generally carry Solar Hijri years while English and Arabic references carry Gregorian years. No calendar conversion is applied. Summary statistics Table 7 summarizes the key statistics of the released dataset.Table 7Summary statistics of the released dataset.Table 7 dummy alt textCategoryCountTotal annotated articles635– Persian352– English142– Arabic141Articles with annotated reference sections621Articles with header annotations only14Total annotated reference strings17,832– Persian reference strings8476– Arabic reference strings5741– English reference strings3615Header metadata fields per article7Reference metadata fields per reference string12 (plus the raw reference string)Number of source conferences and journals>580Number of annotators10Annotation period>12 months Table 7Summary statistics of the released dataset.Table 7 dummy alt textCategoryCountTotal annotated articles635– Persian352– English142– Arabic141Articles with annotated reference sections621Articles with header annotations only14Total annotated reference strings17,832– Persian reference strings8476– Arabic reference strings5741– English reference strings3615Header metadata fields per article7Reference metadata fields per reference string12 (plus the raw reference string)Number of source conferences and journals>580Number of annotators10Annotation period>12 months Summary statistics of the released dataset. Note on the original source documents The annotations were produced from open-access PDF articles. The PDF files themselves are not redistributed in the repository, because their copyright is held by the respective publishers. The dataset is nevertheless self-contained: every annotated text span is stored verbatim inside the JSON files, so no part of the data requires access to the original PDFs in order to be used. Experimental Design, Materials and Methods The dataset was produced through the four-phase procedure shown in Fig. 3.Fig. 3The four-phase procedure by which the dataset was produced. Files rejected in the review phase were returned for re-annotation.Fig 3 dummy alt text The four-phase procedure by which the dataset was produced. Files rejected in the review phase were returned for re-annotation. Phase 1: collection of scientific articles Open-access scientific articles in Persian, Arabic and English were collected from public online repositories and conference proceedings. Source selection prioritized breadth: one of the repositories used contains articles from >580 registered conferences alone. Collecting from many venues rather than a few ensures that the resulting data cover a wide variety of page layouts, formatting styles, fonts, citation formats and publisher conventions, and that they span several scientific disciplines. Phase 2: classification and validation Every collected file was inspected to confirm that it is a scientific article rather than, for example, a table of contents, a proceedings front matter or a poster abstract. Files that passed this check were classified by source venue, by language (Persian, Arabic or English) and by document structure. Language was assigned with Apache Tika [18] and confirmed manually. This phase acted as a quality filter and produced a validated corpus ready for annotation. Phase 3: annotation Annotation was carried out by a team of ten trained annotators over a period of more than twelve months. A Java labelling tool was developed specifically for this task; PDF text and layout were obtained with Apache PDFBox [19]. The tool provides:•Rendering of the PDF article with its full text;•A color-coded tagging interface in which the annotator selects a text span and assigns a metadata label, each label having its own color;•Support for right-to-left (Persian and Arabic) and left-to-right (English) text, including mixed-script reference strings;•Handling of multi-line metadata fields, footnotes and mixed-language content;•Export of the completed annotation to a structured JSON file. •Rendering of the PDF article with its full text; Rendering of the PDF article with its full text; •A color-coded tagging interface in which the annotator selects a text span and assigns a metadata label, each label having its own color; A color-coded tagging interface in which the annotator selects a text span and assigns a metadata label, each label having its own color; •Support for right-to-left (Persian and Arabic) and left-to-right (English) text, including mixed-script reference strings; Support for right-to-left (Persian and Arabic) and left-to-right (English) text, including mixed-script reference strings; •Handling of multi-line metadata fields, footnotes and mixed-language content; Handling of multi-line metadata fields, footnotes and mixed-language content; •Export of the completed annotation to a structured JSON file. Export of the completed annotation to a structured JSON file. Each article was annotated for both header metadata and reference metadata. Annotators worked from a written set of guidelines that fixes the label inventory, the span boundaries to be used for each label and the rules to apply to ambiguous cases, such as reference strings that carry both a journal name and a publisher, or affiliations that run across several lines. Annotation was carried out under supervision, and questions arising during annotation were resolved by the supervisor and, where they affected the conventions themselves, recorded in the guidelines. The exported JSON file for each article is self-contained: it stores the annotated text itself rather than character offsets into the PDF, so neither the PDF nor the labelling tool is needed in order to use the data for training or evaluation. Phase 4: review and quality control Every JSON file produced in Phase 3 was reviewed systematically. Each file was compared field by field against the original PDF in order to verify that the annotated spans are correct and complete. Where discrepancies were found, the article was returned to Phase 3 for re-annotation and re-entered the review queue. Only files that passed review were included in the released dataset. Annotation reliability was assessed on a random 20 % sample of the corpus; the results are reported in the Quality assurance subsection and summarized in Table 8.Table 8Inter-annotator agreement on a random sample of 127 articles (2881 matched reference strings).Table 8 dummy alt textFieldExact match (%)Mean similarityNHeader fieldstitle99.20.992125abstract99.20.998125keywords96.60.990118venue_information16.70.621120doi54.50.54511authors (F1)0.967-127affiliations (F1)0.527-127Reference fieldsreference_title94.30.9832842reference_year96.70.9822812reference_type97.60.9762881reference_journal92.50.9651632reference_publisher85.50.916731reference_volume82.70.8561399reference_pages76.00.9461514reference_number75.80.782894reference_organization68.30.735331reference_doi41.30.430341reference_book_id18.50.18554reference_authors (F1)0.7262881Categorical agreementreference_type(Cohen's Kappa)0.9542881 Table 8Inter-annotator agreement on a random sample of 127 articles (2881 matched reference strings).Table 8 dummy alt textFieldExact match (%)Mean similarityNHeader fieldstitle99.20.992125abstract99.20.998125keywords96.60.990118venue_information16.70.621120doi54.50.54511authors (F1)0.967-127affiliations (F1)0.527-127Reference fieldsreference_title94.30.9832842reference_year96.70.9822812reference_type97.60.9762881reference_journal92.50.9651632reference_publisher85.50.916731reference_volume82.70.8561399reference_pages76.00.9461514reference_number75.80.782894reference_organization68.30.735331reference_doi41.30.430341reference_book_id18.50.18554reference_authors (F1)0.7262881Categorical agreementreference_type(Cohen's Kappa)0.9542881 Inter-annotator agreement on a random sample of 127 articles (2881 matched reference strings). Quality assurance To provide an empirical measure of annotation reliability, an inter-annotator agreement study was conducted on a random sample of 127 articles, corresponding to approximately 20 % of the corpus. The results are summarized in Table 8. Agreement is consistently high for the core bibliographic fields: article title (99.2 % exact match), abstract (99.2 %), keywords (96.6 %), author names (F1 = 0.967), reference title (94.3 %), publication year (96.7 %) and reference type (97.6 %, Cohen's Kappa = 0.954, indicating almost perfect agreement on the scale of Landis and Koch [20]). The reference matching rate between the two annotators was 99.2 %, with 2881 of 2904 reference strings matched. The high agreement on the high-volume fields confirms that the first-pass annotations were of high quality and that the review procedure functioned as a targeted correction mechanism rather than as a wholesale re-annotation. Lower agreement is observed on fields whose boundaries are inherently ambiguous or whose occurrence is sparse. The venue_information field (16.7 % exact match) combines journal name, volume, issue and year into a single free-text string, and the two annotators differed in how they segmented these components. Affiliations (F1 = 0.527) are multi-line and heterogeneous in structure, which makes exact set comparison sensitive to minor formatting differences. DOI appears in only 11 of the 127 sampled articles, so the 54.5 % figure is based on a very small count. Among reference fields, book identifiers (18.5 %) and DOIs (41.3 %) are similarly sparse, and the reference-authors field (F1 = 0.726) reflects variation in how multi-author strings are split into individual names. These patterns are consistent with findings in the metadata extraction literature, where structural fields such as title and year are easier to delineate than composite or optional fields [16]. Tools and software Table 9 lists the software tools used at each stage of the annotation pipeline.Table 9Software used to produce the dataset.Table 9 dummy alt textTool or softwarePurposeCustom Java labelling toolAnnotation interface and JSON exportApache Tika [18]Language detection during classificationApache PDFBox [19]PDF-to-text and PDF-to-HTML conversion in support of annotationJSON / JSON SchemaAnnotation storage format and its formal specification Table 9Software used to produce the dataset.Table 9 dummy alt textTool or softwarePurposeCustom Java labelling toolAnnotation interface and JSON exportApache Tika [18]Language detection during classificationApache PDFBox [19]PDF-to-text and PDF-to-HTML conversion in support of annotationJSON / JSON SchemaAnnotation storage format and its formal specification Software used to produce the dataset. Limitations The dataset covers born-digital PDF articles only; scanned and image-based documents are not represented, and the annotations therefore cannot be used to evaluate extraction from OCR output. The original PDF files are not redistributed because of publisher copyright, so only the structured JSON annotations are provided. Several fields are sparsely populated and should not be used on their own to train a field-specific model. DOI occurs in 15.4 % of header annotations and in 12.1 % of reference strings, reflecting how infrequently DOIs appear in the source articles, particularly in Persian and Arabic publications. Book identifiers occur in 2.3 % of reference strings, and organisation in 12.5 %. Sixty-five reference strings (0.4 %) carry an unknown reference type, and 14 articles (2.2 %) have no reference annotations. The language distribution is unbalanced by design at the article level, with Persian articles making up 55.4 % of the corpus, and the distribution of reference strings differs again from the distribution of articles. The inter-annotator agreement study (Table 8) confirms that agreement is lower on composite fields such as venue_information and on sparse fields such as DOI and book identifiers; users who require these fields should treat them with caution. Only the header and the reference section are annotated. Body text, section structure, figures, tables, equations and captions are not annotated, so the data do not support full-text structure recognition. Annotation was performed on open-access articles only, so paywalled publications, and any formatting conventions specific to them, are not represented. Persian publication years are recorded in the calendar used by the source article and are not normalised. CRediT Author Statement Arash Yousefi Jordehi: Conceptualization, Methodology, Software, Formal analysis, Investigation, Visualization, Writing – review & editing. Reza Fallah: Conceptualization, Methodology, Software, Investigation, Data curation, Visualization, Writing – original draft, Writing – review & editing. Ebrahim Nikmard: Conceptualization, Methodology, Software, Formal analysis, Visualization, Writing – original draft, Writing – review & editing. Mahsa Hosseini Khasheh Heyran: Conceptualization, Methodology, Software, Validation, Investigation, Data curation, Visualization. Seyed Abolghasem Mirroshandel: Resources, Supervision, Project administration, Visualization, Writing – review & editing. Ethics Statement The authors have read and follow the ethical requirements for publication in Data in Brief and confirm that the present work does not involve human subjects, animal experiments or data collected from social media platforms. All articles from which the annotations were derived were obtained from open-access sources. The dataset contains bibliographic metadata only and no personal or sensitive information beyond the author names, affiliations and contact details that already appear in the published articles themselves. The original PDF files are not redistributed, so no publisher copyright is transferred by the deposit. Data Availability The dataset described in this article is publicly available on Mendeley Data under DOI 10.17632/bk52kn69bh.1 [1], without registration or prior request.