CGIAR/agrifood-research-corpus

Agrifood Research Corpus

An open corpus of 40,479 agrifood research documents prepared for retrieval-augmented generation, search, summarization, and language-model development. It holds 34,577 full-text documents and 5,902 title-and-abstract records, each with harmonized metadata (license, resource type, year, geography, and crop and livestock tags from controlled vocabularies).

The corpus is organized by contributing organization. This release contains research outputs from CGIAR. Data from other organizations will be added as separate top-level folders, so you can load the whole corpus or a single organization's subset.

At a glance

Measure	Full text	Abstract only	Total
Records	34,577	5,902	40,479
Tokens (cl100k_base)	381,959,738	2,535,645	384,495,383
Characters	1,705,710,507	12,600,015	1,718,310,522

Organizations: CGIAR · Languages: English (38,068), Spanish (1,567), French (792), Portuguese (52) · Format: Parquet only (14 files, 664 MB).

Repository layout

agrifood-research-corpus/
├── cgiar/
│   ├── full-text/
│   │   ├── permissive/        train-*.parquet
│   │   ├── sharealike/
│   │   ├── noncommercial/
│   │   ├── no-derivatives/
│   │   └── other-licenses/
│   └── abstract-only/
│       └── (same license folders)
└── vocabularies/
    ├── SCiO_Crops.csv
    └── SCiO_Livestock.csv

• Organization (cgiar/, …): who contributed the records.
• Text type: full-text holds the complete text extracted from the document. abstract-only holds the title, a blank line, then the abstract, for records whose full text is not available.
• License category: groups records by the reuse terms of their declared license (see Licensing).

Languages are not split into folders. Use the language column.

Loading

from datasets import load_dataset

ds = load_dataset("CGIAR/agrifood-research-corpus", split="train")            # all organizations
ds = load_dataset("CGIAR/agrifood-research-corpus", "cgiar", split="train")   # CGIAR only

# One text type / license category: point data_files at its folder
ft_permissive = load_dataset("CGIAR/agrifood-research-corpus", data_files="cgiar/full-text/permissive/*.parquet", split="train")

# Filter by language or license with the columns
spanish = ds.filter(lambda r: r["language"] == "es")
commercial_ok = ds.filter(lambda r: r["license_category"] in ("permissive", "sharealike"))

Large full-text partitions can be streamed with streaming=True.

Licensing

There is no single license for this corpus. Every record carries the license declared by its source repository in the rights column, and users must comply with the license of each record they use. Records with unknown, missing, or copyright-protected terms are not included.

Folder / license_category	Licenses	Reuse terms	Full text	Abstract only	Total
permissive	CC0, CC-BY	Attribution only (CC0: none). Commercial use and adaptation allowed.	23,163	4,763	27,926
sharealike	CC-BY-SA	Adaptations must be shared under the same license.	1,616	14	1,630
noncommercial	CC-BY-NC, CC-BY-NC-SA	No commercial use. CC-BY-NC-SA adaptations must also be shared alike.	7,358	348	7,706
no-derivatives	CC-BY-ND, CC-BY-NC-ND	Adapted material may not be shared; CC-BY-NC-ND also bars commercial use.	2,419	777	3,196
other-licenses	Other declared licenses (e.g. GPL, MIT, OGL-UK, Unlicense)	Review each record's license.	21	0	21

Records are grouped by their most restrictive term: CC-BY-NC-SA is under noncommercial and CC-BY-NC-ND under no-derivatives. Check rights for the exact license. No-derivatives licenses deserve particular care. Whether text extraction, chunking, translation, or model training counts as an adaptation depends on the use and the jurisdiction.

rights	Full text	Abstract only	Total
CC-BY	23,116	4,655	27,771
CC-BY-NC	6,034	286	6,320
CC-BY-NC-ND	2,102	768	2,870
CC-BY-SA	1,616	14	1,630
CC-BY-NC-SA	1,324	62	1,386
CC-BY-ND	317	9	326
CC0	47	108	155
UNLICENSE	15	0	15
OGL-UK	3	0	3
GPL	2	0	2
MIT	1	0	1

Content processing

content is Markdown text extracted from the source PDF/DOCX files with Docling, PyPDFium2, and EasyOCR. For this release:

• Figures and tables are not included; their captions are. Markdown tables were removed because many were extracted incompletely (scattered numbers without structure). Text fragments from inside figures (axis labels, diagram boxes) were removed where they appear as runs of short lines next to a figure or table caption. Caption lines (Table 3. …, Figure 2 …) are kept.
• The metadata title is not prepended to full-text content.
• OCR and layout extraction are imperfect. Expect hyphenation artifacts, running headers, and occasional figure text that the heuristics missed.

Schema

Field	Type	Description
gardian_id	string	Unique identifier
organization	string	Contributing organization (cgiar, …)
text_type	string	full_text or abstract_only
language	string	ISO 639-1 code assigned by the GAIA pipeline (en, es, fr, pt); see Limitations
source	string	Asset URL
resource_url	string	Original source/resource URL
title	string	Preferred title
ai_en_title	string	AI-assisted English title, if provided
description	string	Preferred description/abstract
ai_en_description	string	AI-assisted English description, if provided
resource_type	string	Harmonized resource type
release_year	string	Publication year when known
rights	string	Declared license of the record
license_category	string	permissive, sharealike, noncommercial, no-derivatives, other-licenses
crops	struct {values: list[str], categories: list[str]}	SCiO crop vocabulary terms
livestock	struct {values: list[str], categories: list[str]}	SCiO livestock vocabulary terms
keywords	list[str]	Other controlled/uncontrolled terms
geography	struct {regions: list[str], countries: list[str]}	Geographic tags
content	string	Full text (Markdown), or title + abstract

Controlled vocabularies: finding crops and livestock

vocabularies/ holds the SCiO lookup tables used to tag records:

• SCiO_Crops.csv: 180 English crop labels in 10 categories (columns EnglishLabel, Category).
• SCiO_Livestock.csv: 100 English livestock labels in 14 categories (columns English label, Category).

In the data, crops and livestock are structs with independent values and categories lists. The two lists are not positional pairs, so filter on either one independently.

Vocabulary categories

Crop category	Labels
beverage, sugar & industrial crops	10
cereals & pseudocereals	18
forage crops	5
fruits	38
herbs & spices	27
nuts	9
oilseeds & oil crops	11
pulses & grain legumes	13
roots & tubers	8
vegetables	41

Livestock category	Labels
bovines	17
camelids	6
equids	11
farmed deer	6
farmed ratites	3
livestock	1
managed insects	4
pigs	6
poultry	22
poultry and game birds	3
ruminants	1
small livestock	5
small ruminants	12
working animals	3

import csv
from datasets import load_dataset
from huggingface_hub import hf_hub_download

ds = load_dataset("CGIAR/agrifood-research-corpus", "cgiar", split="train", streaming=True)

# All documents tagged with any cereal, using the published vocabulary
vocab = hf_hub_download("CGIAR/agrifood-research-corpus", "vocabularies/SCiO_Crops.csv", repo_type="dataset")
with open(vocab, encoding="utf-8-sig", newline="") as f:
    cereals = {r["EnglishLabel"].casefold() for r in csv.DictReader(f)
               if r["Category"] == "cereals & pseudocereals"}
for row in ds:
    if {x.casefold() for x in (row["crops"] or {}).get("values", [])} & cereals:
        print(row["gardian_id"], row["title"])
        break

SQL over the Parquet files with DuckDB:

import duckdb
duckdb.sql("""
    SELECT gardian_id, title, rights
    FROM 'hf://datasets/CGIAR/agrifood-research-corpus/cgiar/full-text/*/*.parquet'
    WHERE list_contains(livestock.categories, 'small ruminants')
      AND language = 'en'
    LIMIT 20
""").show()

Composition

Languages

Language	Full text	Abstract only	Total
English (en)	32,212	5,856	38,068
Spanish (es)	1,540	27	1,567
French (fr)	779	13	792
Portuguese (pt)	46	6	52

Resource types

Resource type	Full text	Abstract only	Total
Scientific Publication	11,986	4,818	16,804
Report	7,773	42	7,815
Brief	4,501	12	4,513
Poster / Presentation	4,244	33	4,277
Training Material	1,390	6	1,396
Working Paper	1,214	48	1,262
Factsheet	1,243	0	1,243
Dataset / Tabular	0	823	823
Book / Monograph	711	35	746
Other	439	5	444
Online Sources	203	38	241
Report / Case study	233	6	239

Publication years (most frequent)

Year	Full text	Abstract only	Total
2024	4,884	538	5,422
2025	4,416	349	4,765
2023	3,760	606	4,366
2021	3,037	682	3,719
2022	2,954	593	3,547
2020	2,302	548	2,850
2019	2,048	515	2,563
2026	1,624	304	1,928
2018	1,469	402	1,871
2017	1,503	358	1,861
2016	1,427	261	1,688
2015	959	231	1,190

Countries (most frequent tags)

Country	Full text	Abstract only	Total
Ethiopia	3,482	676	4,158
Kenya	2,922	627	3,549
India	2,057	470	2,527
Nigeria	1,746	211	1,957
Uganda	1,534	254	1,788
Tanzania	1,422	233	1,655
Bangladesh	1,285	325	1,610
Ghana	1,194	194	1,388
Colombia	1,247	69	1,316
Viet Nam	1,058	238	1,296
Zambia	982	106	1,088
Malawi	857	134	991

Crops and livestock (most frequent tags)

Crop	Full text	Abstract only	Total
rice	1,900	768	2,668
maize	1,758	388	2,146
wheat	1,572	178	1,750
potato	687	236	923
cassava	799	107	906
banana	645	33	678
sweet potato	435	158	593
barley	539	31	570
chickpea	429	42	471
common bean	418	46	464
cocoa	429	24	453
sorghum	320	103	423

Livestock	Full text	Abstract only	Total
livestock	5,658	951	6,609
cattle	689	338	1,027
sheep	671	159	830
poultry	401	180	581
pig	384	132	516
ruminant	306	58	364
chicken	203	118	321
goat	229	56	285
dairy cattle	137	50	187
cow	82	53	135
pigeon	51	28	79
american bison	37	30	67

Preparation and limitations

Records come from harmonized GARDIAN metadata, document downloads, Docling/PyPDFium2/EasyOCR processing, and GAIA publication preparation. They were selected through GARDIAN/GAIA access and licensing classifications. Downstream users remain responsible for determining what uses are allowed.

• OCR is imperfect, source metadata can be incomplete, multilingual metadata may contain fallback translations, and some records may share publication identities.
• The language label does not always match the language of content. A stopword check suggests about 1–2% of records differ, mostly English text labelled fr or es (e.g. roughly one in six French-labelled full texts is in English).
• Geography, crop, and livestock tags are assigned automatically and should be treated as search aids, not ground truth.
• The corpus is not an evaluation benchmark. The train split is a distribution convention and does not imply that records are cleared for model training (see Licensing).

Credit

Prepared as part of Generative AI for Agriculture (GAIA), funded by the Gates Foundation and UK International Development (FCDO), using content metadata from GARDIAN and technical contributions from SCiO, with CGIAR/IFPRI collaboration. Research outputs remain attributable to their original creators and repositories. Please cite the original work in addition to this corpus.

@misc{agrifood_research_corpus,
  title  = {Agrifood Research Corpus},
  author = {CGIAR Generative AI for Agriculture (GAIA) project},
  year   = {2026},
  url    = {https://huggingface.co/datasets/CGIAR/agrifood-research-corpus}
}