CGIAR/cirad-ai-documents

GAIA Agricultural Research Corpus — CIRAD (English)

A curated, machine-readable corpus of 4,209 agricultural research publications sourced from CIRAD (the French Agricultural Research Centre for International Development), produced by the Generative AI for Agriculture (GAIA) project. Documents are indexed through GARDIAN — CGIAR's agri-food research index — and converted from PDF to structured JSON via the GAIA-CIGI pipeline using GROBID.

This dataset is intended for retrieval-augmented generation, summarization, and fine-tuning of agricultural advisory LLMs.

At a glance

Metric	Value
Documents	4,209
Tokens (re-counted with cl100k_base)	1,463,377
Tokens (precomputed in tokenCount field)	999,716
Total content characters	6,749,631
Mean / median tokens per doc (cl100k_base)	348 / 309
Mean / median content chars per doc	1,604 / 1,463
Language	English (declared at the repo level; no per-doc language metadata in this slice)
On-disk size	14.8 MB (uncompressed JSON)
File count	4,209 JSON files in data/part_1/

Token counts differ between the precomputed tokenCount field and our re-count because the upstream pipeline used a different tokenizer than cl100k_base (the GPT-4 family tokenizer most current LLM consumers see).

Data provenance

All documents share metadata.source = "gardian_index". Publisher distribution:

Host	Docs	Share
agritrop.cirad.fr (CIRAD's open-access repository)	4,200	99.8%
cgspace.cgiar.org	9	0.2%

The 9 documents from cgspace.cgiar.org are the only ones in this dataset that carry populated pagecount, keywords, images, and tables — the rest were extracted with metadata-only output (content + minimal metadata). See Known limitations below.

Splits and file layout

Single train split, single shard:

data/
└── part_1/     4,209 JSON docs   1.46M tokens   14.8 MB

Each file is named <sieverID>.json and contains one document.

Document schema

Every document is a single JSON object. This dataset has a narrower metadata footprint than other GAIA datasets — only five metadata sub-fields are populated, and the optional keywords, images, tables arrays carry data only for the 9 cgspace-sourced documents.

Top-level fields

Field	Type	Always present?	Notes
metadata	object	yes	See Metadata sub-fields below
content	string	yes	Full extracted text (GROBID + PDFBox). Median 1,463 chars, max 61,655
sieverID	string	yes	Internal document identifier (also the filename stem)
pagecount	string	yes	"0" for 4,200 docs; real page counts for 9 docs (range 7–70)
tokenCount	string	yes	Precomputed token count from the original pipeline
keywords	list[string] or null	9 docs have a list	Topical keywords; populated only for cgspace-sourced docs
images	list[string] or null	9 docs have a list	Image keys; fetch at https://cigi-images.s3.us-east-2.amazonaws.com/{key}
tables	list[string] or null	9 docs have a list	Table keys; fetch at https://cigi-tables.s3.us-east-2.amazonaws.com/{key}

Metadata sub-fields

Field	Type	Notes
gardian_id	string	Document identifier within GARDIAN
id	string	Document ID hashed from the source URL
url	string	Source URL (CIRAD's Agritrop repository for 99.8% of docs)
description	string	Abstract or document description
source	string	Always gardian_index

This dataset does not populate the richer metadata fields (title, language, release_year, resource_type, rights, geography, keywords) that some sibling GAIA datasets carry. If those are needed, see e.g. CGIAR/usda-nal-ai-documents-en.

Pipeline

GARDIAN index → PDF fetch (Agritrop / CGSpace)
              → GROBID  (structured text extraction)
              → PDFBox / Tabula  (images & tables — populated for 9 docs)
              → JSON serialization (one file per document)

See the pipeline architecture documentation for full detail.

Loading

from datasets import load_dataset

ds = load_dataset("CGIAR/cirad-ai-documents", split="train")
print(ds)
print(ds[0]["metadata"]["description"][:300])
print(ds[0]["content"][:500])

Known limitations

• Sparse metadata. This dataset has only the 5 metadata sub-fields listed above. There is no title, language, release_year, resource_type, rights, or geography field at the document level. The repo-level language: en tag reflects the slice curation rather than per-doc detection.
• pagecount is unreliable. 4,200 of 4,209 docs have pagecount = "0"; only the 9 cgspace-sourced docs carry a real count.
• License is repo-declared, not per-document. Unlike some sibling GAIA datasets there is no per-document rights field. Treat the repo-level cc-by-4.0 declaration as the curation license; verify the source publisher's license at metadata.url before redistributing individual documents.
• Token counts depend on tokenizer. This card reports cl100k_base (GPT-4 family) counts as the headline number; the precomputed tokenCount field uses a different (older) tokenizer.

Citation

@misc{cgiar_gaia_cirad_en,
  title  = {GAIA Agricultural Research Corpus — CIRAD slice (English)},
  author = {CGIAR Generative AI for Agriculture (GAIA) project},
  year   = {2025},
  doi    = {10.57967/hf/7746},
  url    = {https://huggingface.co/datasets/CGIAR/cirad-ai-documents}
}

Acknowledgements

This dataset was developed for the Generative AI for Agriculture (GAIA) project, funded by the Bill & Melinda Gates Foundation and UK International Development (FCDO), in collaboration between CGIAR, CIRAD, and SCiO.