EVE-Corpus

Paper | GitHub

EVE-Corpus is a large-scale, cleaned, and anonymized text corpus of Earth Observation (EO) documents formatted in Markdown.
It is designed to support research in EO and domain-specific LLM training.

The corpus contains 186k Markdown files, sourced from peer-reviewed journals, EO websites and scientific repositories.

---

Dataset Features

• file_path (string): the file path of the source document within the s3 bucket.
• text (string): full document content in Markdown format.
• publisher (string): Publisher or source of the document.
• title (string): Title of the document.
• authors (string): Author(s) of the document.
• url (string): Source URL.
• year (int64): Year of publication.
• license (string): Document license information.

Dataset Summary

EVE-Corpus aggregates high-quality EO-related text covering topics such as:

• Satellite missions
• Remote sensing techniques
• Climate and atmospheric science
• Land, ocean, and cryosphere monitoring
• Environmental modelling and geospatial analytics

All documents have been deduplicated, cleaned, and anonymized to remove personal information.

---

Data Distribution

The dataset includes 2.8B tokens from 22 EO-related sources.

publisher	tokens	percentage
MDPI	1.3B	46.55
Copernicus	723M	25.89
NCBI	485M	17.37
ISPRS	74.1M	2.65
Wiley	71.6M	2.56
Elsevier	43.7M	1.56
Cambridge Press	40.3M	1.44
Springer	25.4M	0.91
Taylor and Francis	11.8M	0.42
AMS	5.7M	0.2
SAGE	3.8M	0.14
NASA	2.5M	0.09
Arxiv	2.1M	0.08
IEEE	992K	0.04
EGUP	671K	0.02
Oxford Academic	576K	0.02
Iopscience	507K	0.02
Frontiers	307K	0.01
EOGE	245K	0.01
EOS	96K	0.003
MIT	83K	0.003
Uk Met Office	7K	0.0002
Total	2.8B	100

---

Source Collection

The dataset was constructed using this scraping pipeline.

---

Preprocessing Pipeline

The dataset was processed using the following pipeline:

• Extraction

  • Supports PDF, HTML, XML, Markdown, and nested folder structures.
  • Automatically detects file formats unless explicitly specified.

• Deduplication

  • Performs exact matching using SHA-256 checksums.
  • Supports LSH-based near-duplicate detection (configurable: shingle size, permutations, similarity threshold).

• Cleaning

  • Removes irregularities and noise artifacts.
  • Corrects LaTeX equations and tables using LLM assistance.

• PII Removal

  • Automatically masks names and emails using the Presidio framework.

• Metadata Extraction

  • Extracts Title, Authors, DOI, URL, Year, Journal, and Citation Count from scientific papers.

• Export

  • Saves processed content in multiple formats (default: Markdown).

---

Attribution info

Attributions to the original authors can be found in the attribution.txt. Each dataset entry has the file_path column that maps the document to the corresponding attribution entry

Citation

If you use this project in academic or research settings, please cite:

@misc{atrio2026evedomainspecificllmframework,
      title={{EVE}: A Domain-Specific {LLM} Framework for Earth Intelligence}, 
      author={Àlex R. Atrio and Antonio Lopez and Jino Rohit and Yassine El Ouahidi and Marcello Politi and Vijayasri Iyer and Umar Jamil and Sébastien Bratières and Nicolas Longépé},
      year={2026},
      eprint={2604.13071},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2604.13071}, 
}