CGIAR/gardian-cigi-ai-documents

GAIA / GARDIAN-CIGI Agricultural Research Corpus (English)

A curated, machine-readable corpus of 65,550 agricultural research publications drawn from across the CGIAR centers and 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 and supporting extractors.

At a glance

Metric	Value
Documents	65,550
Pages (total)	1,780,047 (across 55,780 docs with populated pagecount)
Tokens (re-counted with cl100k_base)	474,717,562
Tokens (precomputed in tokenCount field)	343,498,673
Total content characters	2,249,559,219
Mean / median tokens per doc (cl100k_base)	7,242 / 2,378
Mean / median content chars per doc	34,318 / 11,537
Language	English (declared at the repo level; no per-doc language metadata)
On-disk size	1.32 GB (Parquet) / 2.36 GB (raw JSON shards)
File count	1 Parquet shard at data/train.parquet; 65,550 JSON files at data/part_{1..9}/ (mirror of the same content)

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). The cl100k_base re-count is ~38% higher because that tokenizer splits English text into smaller pieces than the older WordPiece-style tokenizers.

Document length is highly skewed. The corpus contains both short abstracts and book-length reports — p99 is 80,462 tokens and the longest single document is 1,150,715 tokens (3.1M chars). Plan chunking accordingly for RAG use.

Data provenance

All documents share metadata.source = "gardian_index". Publisher distribution (top URL hosts):

Host	Docs	Share
cgspace.cgiar.org (CGIAR's central repository)	51,929	79.2%
oar.icrisat.org (ICRISAT Open Access Repository)	8,057	12.3%
digitalarchive.worldfishcenter.org (WorldFishCenter)	2,271	3.5%
www.worldagroforestry.org (ICRAF)	945	1.4%
dataverse.harvard.edu	347	0.5%
www.cifor.org (CIFOR)	300	0.5%
ciat-library.ciat.cgiar.org (CIAT)	125	0.2%
Other CGIAR centers and external hosts	1,576	2.4%

The corpus reflects research output from across the CGIAR network of centers (CGSpace alone hosts most of CGIAR's institutional output) plus selected external publications discovered via the GARDIAN index.

Splits and file layout

Single train split. The repository ships two equivalent layouts:

data/
├── train.parquet         65,550 rows   1.32 GB   <-- default loader path
├── part_1/   8,016 JSON docs   58.2M tokens
├── part_2/   8,030 JSON docs   57.0M tokens
├── part_3/   8,008 JSON docs   58.6M tokens
├── part_4/   8,030 JSON docs   59.4M tokens
├── part_5/   8,002 JSON docs   58.4M tokens
├── part_6/   7,984 JSON docs   55.0M tokens
├── part_7/   7,992 JSON docs   59.3M tokens
├── part_8/   7,975 JSON docs   57.6M tokens
└── part_9/   1,513 JSON docs   11.3M tokens

(Token counts are cl100k_base.) The Parquet file is the canonical copy used by load_dataset() and the HF dataset viewer. The JSON shards under data/part_{1..9}/ are kept as a per-document raw mirror for users who want individual <sieverID>.json files.

Document schema

Every document is a single JSON object with the structure below. About 85% of documents are "fully extracted" (with populated keywords, images, tables, and a real pagecount); the remaining ~15% are "text-only" — content is populated but the derived fields are null.

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 11,537 chars, max 3.12M chars
sieverID	string	yes	Internal document identifier (also the filename stem)
pagecount	string	yes	Numeric string. Populated (>"0") for 55,780 / 65,550 docs (85%); total 1,780,047 pages
tokenCount	string	yes	Precomputed token count from the original pipeline
keywords	list[string] or null	55,780 docs have a list	Topical keywords from the source index
images	list[string] or null	55,861 docs have a list	Image keys; fetch at https://cigi-images.s3.us-east-2.amazonaws.com/{key} (4.09M keys total)
tables	list[string] or null	55,870 docs have a list	Table keys; fetch at https://cigi-tables.s3.us-east-2.amazonaws.com/{key} (1.80M keys total)

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 (typically a CGSpace bitstreams/<uuid>/retrieve endpoint)
description	string	Abstract or document description
source	string	Always gardian_index

The GARDIAN-CIGI corpus has a narrow metadata footprint (5 sub-fields). The sibling usda-nal-ai-documents-en slice carries richer per-document metadata (title, language, release_year, resource_type, rights, geography) — that enrichment was applied per-slice rather than to the parent corpus.

Pipeline

GARDIAN index → PDF fetch (CGSpace, ICRISAT OAR, WorldFishCenter,
                            ICRAF, CIFOR, CIAT, Harvard Dataverse, …)
              → GROBID  (structured text extraction, document body)
              → PDFBox  (image extraction)
              → Tabula  (table extraction)
              → JSON serialization (one file per document)
              → semantic-coherence chunking applied downstream by consumers

See the pipeline architecture documentation and chunking method notes for full detail.

Loading

from datasets import load_dataset

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

The dataset is gated — accept the terms on the dataset page and pass your HF token (HF_TOKEN env var or huggingface-cli login) when loading.

Streaming is recommended at this size (~1.3 GB Parquet, 65k docs):

ds = load_dataset("CGIAR/gardian-cigi-ai-documents", split="train", streaming=True)
for doc in ds:
    ...

Known limitations

• Sparse metadata. Only 5 metadata sub-fields are populated. There is no per-document title, language, release_year, resource_type, rights, or geography. See usda-nal-ai-documents-en for an example slice with enriched metadata.
• ~15% of documents are text-only. 9,770 of 65,550 docs have null keywords, images, and tables. Their content is still populated; only the derived fields are missing.
• License is repo-declared, not per-document. There is no per-document rights field in this slice. Verify the source publisher's license at metadata.url before redistributing individual documents. CGIAR-published documents on CGSpace are predominantly CC-BY.
• Document length is highly skewed. Length p50 is 2,378 tokens but p99 is 80,462 and max is 1.15M. Chunk before processing for any task that assumes uniform document length.
• 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_gardian_cigi_en,
  title  = {GAIA / GARDIAN-CIGI Agricultural Research Corpus (English)},
  author = {CGIAR Generative AI for Agriculture (GAIA) project},
  year   = {2025},
  doi    = {10.57967/hf/4327},
  url    = {https://huggingface.co/datasets/CGIAR/gardian-cigi-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 and SCiO.