CGIAR/embrapa-ai-documents

GAIA / EMBRAPA Agricultural Documents

A curated, machine-readable corpus of 46,736 short agricultural publications sourced from EMBRAPA — the Brazilian Agricultural Research Corporation. Despite the repository name suffix -en, the corpus is predominantly Portuguese (~70%) with the remaining ~30% in English; see Known limitations below. Produced by the Generative AI for Agriculture (GAIA) project. Documents are indexed through GARDIAN and converted from PDF to structured JSON via the GAIA-CIGI pipeline using GROBID.

Compared to other GAIA datasets, these documents are short — median 1,280 characters / 337 tokens. They are abstracts or short summary records rather than full-text articles, reflecting the structure of EMBRAPA's Alice (institutional digital library) records. It is suited to retrieval-augmented generation, semantic search, and Brazilian-agriculture-focused fine-tuning.

At a glance

Metric	Value
Documents	46,736
Tokens (re-counted with cl100k_base)	17,735,184
Tokens (precomputed in tokenCount field)	10,115,525
Total content characters	66,556,091
Mean / median tokens per doc (cl100k_base)	379 / 337
Mean / median content chars per doc	1,424 / 1,280
Max content (chars / tokens)	59,627 / 20,733
Language	Mixed — ~70% Portuguese, ~30% English (sampled). No per-doc language metadata is stored; both languages are declared at the repo level
On-disk size	82 MB (Parquet) / 146 MB (raw JSON shards)
File count	1 Parquet shard at data/train.parquet; 46,736 JSON files at data/part_{1..6}/ (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).

Document length is unusually uniform for the GAIA family. p95 is only 736 tokens (2,708 chars) — this dataset is suited to RAG queries with short retrieval contexts rather than long-document summarization.

Data provenance

All 46,736 documents come from a single source — EMBRAPA's Alice (Acesso Livre à Informação Científica da Embrapa) digital library at www.alice.cnptia.embrapa.br. All documents share metadata.source = "gardian_index".

Splits and file layout

Single train split. The repository ships two equivalent layouts:

data/
├── train.parquet      46,736 rows    82 MB   <-- default loader path
├── part_1/   9,054 JSON docs   3.45M tokens
├── part_2/   9,054 JSON docs   3.40M tokens
├── part_3/   9,054 JSON docs   3.44M tokens
├── part_4/   9,054 JSON docs   3.44M tokens
├── part_5/   9,054 JSON docs   3.46M tokens
└── part_6/   1,466 JSON docs   0.56M 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..6}/ 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. Only the core fields are present — there are no derived images, tables, keywords, or populated pagecount.

Top-level fields

Field	Type	Always present?	Notes
metadata	object	yes	See Metadata sub-fields below
content	string	yes	Full extracted text (typically a short abstract). Median 1,280 chars, max 59,627. ~70% Portuguese, ~30% English
sieverID	string	yes	Internal document identifier (also the filename stem)
pagecount	string	yes	Currently "0" for every document — page counts were not populated for this slice
tokenCount	string	yes	Precomputed token count from the original pipeline
keywords	list[string]	yes (always empty/null in this slice)	Schema reserves a list type for forward compatibility; no values populated
images	list[string]	yes (always empty/null in this slice)	Schema reserves a list type for forward compatibility; no values populated
tables	list[string]	yes (always empty/null in this slice)	Schema reserves a list type for forward compatibility; no values populated

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 (always under www.alice.cnptia.embrapa.br)
description	string	Abstract or document description
source	string	Always gardian_index

The EMBRAPA dataset does not populate the richer metadata fields (title, language, release_year, resource_type, rights, geography) that some sibling GAIA datasets carry.

Pipeline

GARDIAN index → PDF fetch (EMBRAPA Alice repository)
              → GROBID  (structured text extraction)
              → JSON serialization (one file per document)

images, tables, and pagecount are not extracted for this slice — content and description are the payload.

See the pipeline architecture documentation for full detail.

Loading

from datasets import load_dataset

ds = load_dataset("CGIAR/embrapa-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.

The dataset fits comfortably in memory at 113 MB; streaming is optional:

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

Known limitations

• Language mismatch with repo name. Despite the -en suffix in the repo ID, a random 2,000-document sample shows ~70% are written in Portuguese and only ~30% are in English. No per-document language field is stored, so language must be detected at query time if filtering is required. Consumers expecting only-English text should pre-filter with a language-ID library (e.g., langdetect or fasttext).
• Sparse content. Documents are short (median 1,280 chars). Many are likely catalog records or abstracts rather than full-text articles. If full-text agricultural research is the goal, sibling datasets like ifpri-ai-documents or gardian-cigi-ai-documents carry longer documents.
• No derived fields. keywords, images, tables, and pagecount are not populated. Only content and the 5 metadata sub-fields above are available.
• License is repo-declared, not per-document. EMBRAPA publishes most of its output under permissive Brazilian government licenses, but verify 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_embrapa_en,
  title  = {GAIA / EMBRAPA Agricultural Documents (English)},
  author = {CGIAR Generative AI for Agriculture (GAIA) project},
  year   = {2025},
  doi    = {10.57967/hf/7747},
  url    = {https://huggingface.co/datasets/CGIAR/embrapa-ai-documents}
}

Acknowledgements

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