DarjaCore/asjp-cerist

ASJP / CERIST Academic PDF Corpus

Overview

DarjaCore ASJP/CERIST is a large-scale collection of academic PDF documents collected from the Algerian ASJP (Algerian Scientific Journal Platform) ecosystem associated with CERIST (Centre de Recherche sur l'Information Scientifique et Technique).

The dataset preserves the original PDF documents rather than converting them into extracted text. This allows researchers to perform their own OCR, document parsing, layout analysis, multilingual processing, metadata extraction, and other downstream tasks.

The collection contains tens of thousands of academic PDF files covering Algerian scientific and scholarly publications.

Dataset Statistics

Statistic	Value
PDF files	37,807
Metadata records	21,604
Total PDF size	~50.30 GB
Total pages	579,683
Extracted text characters*	1,094,404,357
Extracted text words*	166,828,860
Estimated tokens*	~216.9 million

* Text statistics were obtained during corpus analysis. They should not be interpreted as the exact amount of machine-readable text available in every PDF.

Document Composition

The corpus contains a mixture of digitally generated PDFs and scanned documents.

PDF type	Files	Percentage
Text PDFs	6,231	16.48%
Image PDFs	494	1.31%
Mixed PDFs	30,981	81.95%
Empty / no detected text	50	0.13%
Processing errors	51	—

The large proportion of mixed PDFs means that many documents may contain both machine-readable text and scanned/image-based pages.

Dataset Structure

The repository is organized into numbered subdirectories to avoid excessive numbers of files in a single directory.

asjp-cerist/
├── README.md
├── metadata.jsonl
└── pdfs/
    ├── 000/
    │   ├── 00001.pdf
    │   ├── 00001_sommaire.pdf
    │   └── ...
    ├── 001/
    ├── 002/
    ├── 003/
    ├── 004/
    ├── 005/
    ├── 006/
    └── 007/

Each PDF is identified by a normalized numeric identifier.

Metadata

The metadata.jsonl file contains one JSON object per line.

Available fields include:

• title
• journal_title
• volume
• issue
• first_page
• last_page
• publication_date
• abstract
• keywords
• pdf_url
• sommaire_pdf_url
• article_url
• authors
• authors_emails
• institutions
• local_pdf_path
• local_sommaire_pdf_path

Example:

{
  "title": "...",
  "journal_title": "...",
  "volume": "...",
  "issue": "...",
  "publication_date": "...",
  "abstract": "...",
  "keywords": ["...", "..."],
  "pdf_url": "...",
  "article_url": "...",
  "authors": ["..."],
  "institutions": ["..."],
  "local_pdf_path": "pdfs/000/00001.pdf"
}

Why Raw PDFs?

This dataset intentionally keeps the original documents in PDF format.

This makes the corpus useful for:

• OCR research
• Arabic OCR
• French OCR
• Multilingual document understanding
• PDF parsing
• Document layout analysis
• Academic information extraction
• Metadata extraction
• Citation extraction
• Table and figure extraction
• Retrieval-Augmented Generation (RAG)
• Building academic search engines
• Training document-processing models
• Creating downstream text datasets

Keeping the original PDFs allows researchers to select their own extraction and preprocessing pipeline.

Languages

The corpus contains multilingual academic material, including:

• Arabic
• French
• English

Individual documents may contain more than one language.

Data Processing

The preparation pipeline included:

1. Collecting PDF documents.
2. Collecting associated bibliographic metadata.
3. Normalizing PDF identifiers.
4. Organizing PDFs into subdirectories.
5. Updating metadata paths.
6. Validating the relationship between metadata and PDF files.
7. Converting metadata into JSON Lines format.
8. Validating the final dataset structure.

Final validation confirmed:

Metadata records         : 21,604
Actual PDFs              : 37,807

Main PDFs referenced     : 19,053
Sommaire PDFs referenced : 18,754

Unique referenced PDFs   : 37,807

Missing main PDFs        : 0
Missing sommaire PDFs    : 0
Unreferenced PDFs        : 0

JSONL validation         : OK

Intended Uses

This corpus can be useful for research and development involving:

• Algerian academic literature
• Arabic NLP
• French NLP
• Multilingual NLP
• Information retrieval
• Semantic search
• Retrieval-Augmented Generation
• Academic question answering
• OCR
• Document AI
• Digital libraries
• Knowledge extraction
• Academic language-model datasets

Scanned Documents

A significant portion of the corpus contains scanned or partially scanned pages.

Consequently, standard PDF text extraction may not recover all of the document content. OCR may be required for image-based pages.

A typical processing pipeline could be:

PDF
 │
 ├── PDF text extraction
 │
 ├── OCR for image-based pages
 │
 ├── Language identification
 │
 ├── Text normalization
 │
 ├── Deduplication
 │
 ├── Metadata alignment
 │
 └── Training / Retrieval Dataset

For Arabic documents, researchers should pay particular attention to Unicode normalization, Arabic character variants, OCR errors, diacritics, and document layout.

Main PDF vs. Sommaire PDF

The dataset contains both main article PDFs and documents identified with the _sommaire.pdf suffix.

Researchers building text datasets may wish to treat these as separate document types during preprocessing.

Data Quality

The dataset was validated to ensure that:

• Metadata is valid JSONL.
• Referenced PDF files exist.
• Main PDF paths are valid.
• Sommaire PDF paths are valid.
• No local PDFs were left unreferenced.
• PDF identifiers follow the normalized naming convention.

Citation

If you use this dataset in research, please cite the original ASJP/CERIST sources and this dataset repository.

@dataset{darjacore_asjp_cerist,
  title     = {ASJP / CERIST Academic PDF Corpus},
  author    = {DarjaCore},
  publisher = {Hugging Face},
  year      = {2026},
  url       = {https://huggingface.co/datasets/DarjaCore/asjp-cerist}
}

Acknowledgements

This dataset was assembled by DarjaCore as part of efforts to build large-scale Arabic and Algerian-language resources for research in natural language processing, information retrieval, document understanding, and language model development.

The underlying documents originate from the ASJP/CERIST academic publication ecosystem. Credit belongs to the original authors, journals, publishers, and source platforms.

License and Rights

The PDFs in this repository originate from third-party academic publications.

Users are responsible for determining whether their intended use of individual documents complies with the copyright, licensing, and access conditions applicable to those documents and their respective publishers.

This repository does not grant redistribution rights for the underlying publications.

Before making the dataset publicly accessible or using it for commercial purposes, verify the applicable rights and licenses of the source documents.

Disclaimer

This dataset is provided for research and educational purposes.

DarjaCore does not claim ownership of the underlying academic publications. The original authors, journals, publishers, and rights holders retain their respective rights.

The dataset may contain documents subject to different copyright and licensing conditions.