CGIAR/ifpri-ai-documents

GAIA Agricultural Research Corpus — IFPRI slice (English)

A curated, machine-readable corpus of 12,007 agricultural research and policy documents sourced from the International Food Policy Research Institute (IFPRI), 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.

The slice is the IFPRI-attributed portion of the broader gardian-cigi-ai-documents corpus, focusing on agricultural policy research: working papers, discussion papers, journal articles, technical reports, country studies, and books. It is intended for retrieval-augmented generation, summarization, and fine-tuning of agricultural advisory and policy LLMs.

At a glance

Metric	Value
Documents	12,007
Pages (total)	362,716
Tokens (re-counted with cl100k_base)	97,865,144
Tokens (precomputed in tokenCount field)	75,284,385
Total content characters	495,733,936
Mean / median tokens per doc (cl100k_base)	8,151 / 3,753
Mean / median content chars per doc	41,287 / 19,294
Mean / median pages per doc	30.3 / ~15
Language	English (declared at the repo level; no per-doc language metadata)
On-disk size	267 MB (Parquet) / 528 MB (raw JSON shards)
File count	1 Parquet shard at data/train.parquet; 12,007 JSON files at data/part_{1,2}/ (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).

Compared to other GAIA slices, IFPRI documents are long-form: many are 50–700-page books, working papers, and country reports. The 95th percentile is 26,106 tokens per doc and the longest single document is 423,050 tokens (a 768-page report). Plan chunking accordingly for RAG use.

Data provenance

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

Host	Docs	Share
cgspace.cgiar.org (CGSpace repository)	11,655	97.1%
dataverse.harvard.edu	235	2.0%
www.canr.msu.edu	7	0.1%
www.fao.org	5	<0.1%
borlaug.tamu.edu	4	<0.1%
64 other hosts	101	0.8%

The vast majority of documents come through CGSpace — CGIAR's central content repository, where IFPRI publishes its research output.

Splits and file layout

Single train split. The repository ships two equivalent layouts:

data/
├── train.parquet   12,007 rows   267 MB   <-- default loader path
├── part_1/         8,248 JSON docs   66.3M tokens   341 MB
└── part_2/         3,759 JSON docs   31.6M tokens   163 MB

(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_{1,2}/ 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. The IFPRI slice has a narrow metadata footprint (5 sub-fields) but is the most complete GAIA slice for derived content: pagecount, keywords, images, and tables are populated for ≥99.7% of 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 ~19k chars, max 2.35M chars
sieverID	string	yes	Internal document identifier (also the filename stem)
pagecount	string	yes	Numeric string. Populated (>"0") for 11,971 / 12,007 docs (99.7%). Total 362,716 pages
tokenCount	string	yes	Precomputed token count from the original pipeline
keywords	list[string] or null	11,971 docs have a list	Topical keywords
images	list[string] or null	11,978 docs have a list	Image keys; fetch at https://cigi-images.s3.us-east-2.amazonaws.com/{key} (356,989 keys total)
tables	list[string] or null	11,981 docs have a list	Table keys; fetch at https://cigi-tables.s3.us-east-2.amazonaws.com/{key} (365,565 keys total)

When present, image lists have a median of 4 keys (max 68,603 for one extremely large report); table lists a median of 16 keys (max 1,003).

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 IFPRI slice does not populate title, language, release_year, resource_type, rights, or geography at the metadata level. If those are needed, see e.g. CGIAR/usda-nal-ai-documents-en.

Pipeline

GARDIAN index → PDF fetch (CGSpace / Harvard Dataverse / other)
              → 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/ifpri-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 for the IFPRI slice since the on-disk size is ~528 MB:

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

Known limitations

• Sparse metadata. This slice 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.
• License is repo-declared, not per-document. Unlike some sibling GAIA datasets (e.g., usda-nal-ai-documents-en) there is no per-document rights field. IFPRI's own publications are predominantly CC-BY; verify the source publisher's license at metadata.url for non-IFPRI documents before redistributing.
• Document length is highly skewed. The corpus contains both short briefs (a few hundred tokens) and book-length reports (>400k tokens). Naive batching by document will produce extreme variance — chunk before processing.
• 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_ifpri_en,
  title  = {GAIA Agricultural Research Corpus — IFPRI slice (English)},
  author = {CGIAR Generative AI for Agriculture (GAIA) project},
  year   = {2025},
  doi    = {10.57967/hf/7702},
  url    = {https://huggingface.co/datasets/CGIAR/ifpri-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, IFPRI, and SCiO.