CGIAR/usda-nal-ai-documents-en

GAIA Agricultural Research Corpus — USDA-NAL

A curated, machine-readable corpus of 22,848 agricultural research publications in English, produced by the Generative AI for Agriculture (GAIA) project. The corpus is sourced from USDA National Agricultural Library and converted from PDF into structured JSON via the GAIA-CIGI pipeline using GROBID and supporting extractors. The output is designed for downstream LLM use: retrieval- augmented generation, summarization, question answering, and domain fine-tuning for agricultural advisory applications.

At a glance

Metric	Value
Documents	22,848
Tokens (re-counted with cl100k_base)	37,745,697
Tokens (precomputed in tokenCount field)	21,091,841
Total content characters	141,076,399
Mean / median tokens per doc (cl100k_base)	1,652 / 356
Mean / median content chars per doc	6,175 / 1,710
Language	English (100%)
Resource type	Scientific Publication (100%)
Release-year span	2007–2026 (modal year: 2022)
On-disk size	199 MB (uncompressed JSON)
File count	22,848 JSON files in data/part_{1,2,3}/

The repository name carries the USDA-NAL label because the corpus is the English slice aligned with US Department of Agriculture / National Agricultural Library (NAL) thematic coverage; the underlying aggregation, deduplication, and full-text retrieval all run through GARDIAN. Token counts differ between the precomputed tokenCount field and our re-count because the upstream pipeline used a different tokenizer than cl100k_base.

Data provenance

All documents share metadata.source = "gardian_index". The publisher distribution (top URL hosts) gives a more useful picture:

Publisher / host	Docs	Share
api.elsevier.com	20,430	89.4%
link.springer.com	883	3.9%
ejmanager.com	198	0.9%
bio-conferences.org	170	0.7%
jurnal.uns.ac.id	162	0.7%
cambridge.org	162	0.7%
cgspace.cgiar.org	154	0.7%
85 other hosts	889	3.9%

12,233 documents (54%) carry one or more geographic tags in metadata.geography. Top regions: World (12,233), Americas (6,086), Northern America (5,155), United States of America (5,036), Asia (4,658), Africa (2,685), Sub-Saharan Africa (2,524), Europe (1,724), Western Africa (1,521), Western Asia (1,420).

The corpus skews recent — over 60% of documents were published 2020 or later — and is topically broad with strong concentrations in genomics, aquaculture, climate change, soils and nutrients, and food security (see Topical coverage below).

Splits and file layout

The corpus ships as a single train split of 22,848 examples, physically partitioned into three roughly equal shards for git/LFS hygiene:

data/
├── part_1/     9,046 JSON docs   14.6M tokens   80 MB
├── part_2/     9,039 JSON docs   15.3M tokens   84 MB
└── part_3/     4,763 JSON docs    7.8M tokens   43 MB

(Token counts are cl100k_base.)

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

Document schema

Every document is a single JSON object. Field types reflect what actually appears across the corpus (verified by exhaustive scan):

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 1,710 chars, max 381,501 chars
sieverID	string	yes	Internal document identifier (also used as filename stem)
pagecount	string	yes	Currently "0" for every document — page counts were not populated in this release
tokenCount	string	yes	Precomputed token count from the original pipeline (different tokenizer; sums to ~21.1M)
keywords	null	yes (always null)	Reserved for future use; the populated keyword list lives in metadata.keywords
images	list[string] or null	9.5% have a list	Keys to images extracted from the source PDF. Fetch at https://cigi-images.s3.us-east-2.amazonaws.com/{key}
tables	list[string] or null	9.5% have a list	Keys to tables extracted by Tabula. Fetch at https://cigi-tables.s3.us-east-2.amazonaws.com/{key}

When present, image lists have a median of 4 keys (max 300); table lists a median of 11 keys (max 119). The dataset contains 13,999 image references and 25,827 table references in total.

Metadata sub-fields

Field	Type	Notes
gardian_id	string	Document identifier within GARDIAN
dataNODE_id	string	Internal CGIAR-CIGI identifier
id	string	Document ID hashed from the source URL
url	string	Source URL (typically a publisher full-text endpoint)
title	string	Publication title
description	string	Abstract or document description
language	string	Always English in this slice
release_year	string	Publication year, e.g. "2022"
resource_type	string	Always Scientific Publication in this slice
source	string	Always gardian_index
rights	string	Per-document license — see Licensing below
keywords	list[string]	Topical keywords (mean ~21 per doc, max 142)
geography	list[string]	Geographic tags in the format "{value=<name>, code=UNM49:<code>}" (UN M49 region codes)

Topical coverage

The 73,251 unique terms in metadata.keywords give a clear picture of what the corpus covers. The 15 most common terms across all documents:

Keyword	Docs
models	2,436
agriculture	2,418
genes	1,663
aquaculture	1,552
temperature	1,500
soil	1,447
climate change	1,422
species	1,357
humans	1,284
data collection	1,245
nitrogen	1,221
risk	1,199
corn	1,130
climate	1,069
China	1,048

Pipeline

GARDIAN index → PDF fetch (publisher / OA repository)
              → 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

The dataset is configured for the standard datasets loader:

from datasets import load_dataset

ds = load_dataset("CGIAR/usda-nal-ai-documents-en", split="train")
print(ds)
print(ds[0]["metadata"]["title"])
print(ds[0]["content"][:500])

Streaming is also supported (avoids materializing 199 MB):

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

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.

Licensing

The dataset card declares CC-BY-4.0 at the repository level, but the per-document license varies and is recorded in metadata.rights:

metadata.rights	Docs	Share
CC-BY-NC-ND	8,011	35.1%
CC-BY	5,976	26.2%
CC-BY-NC	710	3.1%
CC0	29	0.1%
CC-BY-NC-SA	8	<0.1%
OGL-UK	6	<0.1%
(no license recorded)	8,108	35.5%

When using individual documents, check metadata.rights and obey the applicable license. Some documents (CC-BY-NC*, CC-BY-NC-ND) restrict commercial use and/or derivative works.

Known limitations

• pagecount is unpopulated ("0" for every document). Total page count cannot be derived from this dataset.
• Top-level keywords is always null — the populated list is in metadata.keywords instead. The two fields are not equivalent.
• metadata.geography is in a structured string format ("{value=..., code=UNM49:...}"), not parsed JSON. Consumers needing programmatic access should split on value= / code=.
• ~35% of documents have no rights metadata. Treat these as unknown-license unless you can verify from the source URL.
• Token counts depend on tokenizer. This card reports cl100k_base (GPT-4 family) counts as the headline number, because that is what most current LLM consumers will encounter. The corpus's own precomputed tokenCount field uses a different (older) tokenizer.

Citation

If you use this dataset, please cite the GAIA project and the dataset DOI:

@misc{cgiar_gaia_usda_nal_en,
  title  = {GAIA Agricultural Research Corpus — English (USDA-NAL slice)},
  author = {CGIAR Generative AI for Agriculture (GAIA) project},
  year   = {2025},
  doi    = {10.57967/hf/9130},
  url    = {https://huggingface.co/datasets/CGIAR/usda-nal-ai-documents-en}
}

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.