AfriSpeech/africa-papers

Africa Papers: The Pan-African Academic Knowledge Corpus

Africa Papers is a massive, multi-domain pre-training corpus of peer-reviewed research papers, preprints, agricultural manuals, and scientific literature produced in and about Africa. Curated by AfriSpeech (Author: Mich-Seth Owusu, michseth@afrispeech.org), it provides high-signal African academic knowledge for training and evaluating large language models.

Preamble: Why This Dataset Matters

The Blind Spot in Global Artificial Intelligence

Modern foundation language models (such as LLaMA, GPT-4, and Qwen) are trained on trillions of tokens extracted from the public web and international academic repositories (e.g. Common Crawl, The Pile, S2ORC, and Dolma). However, African scientific, medical, legal, and agricultural scholarship constitutes less than 0.5% of global pre-training corpora.

This severe geographic and intellectual imbalance leads to profound failure modes in state-of-the-art AI systems:

• Biomedical & Epidemiological Blind Spots: Models often fail to accurately reason about tropical diseases (such as malaria variants, sickle cell genetics, lassa fever, or maternal health) within localized African healthcare delivery systems.
• Agricultural Misalignment: Standard AI assistants default to temperate Western industrial monoculture rather than the intercropped, climate-resilient agro-ecologies of African smallholders managing cassava, yams, cowpea, plantains, or sorghum.
• Jurisprudential & Policy Gaps: Fundamental African governance frameworks—including customary legal systems, community land tenure, and regional economic treaties—are largely absent from Western legal datasets.
• Distorted Scientific Representation: Because Western web-scrapes disproportionately reflect Africa through crisis reporting, standard LLMs lack exposure to African scientists as primary intellectual authorities in pure mathematics, astrophysics, robotics, material engineering, and natural language processing.

African Intellectual Sovereignty & Scientific Authority

Africa Papers was engineered to systematically dismantle this structural omission. Rather than relying on external narratives about Africa, this corpus gathers the primary peer-reviewed scholarship, agricultural extension treatises, and STEM preprints produced by African researchers and African institutions across all 54 nations.

By consolidating over 10 Billion tokens of structured, layout-aware academic Markdown, this dataset provides global AI researchers, African research labs, and open-source model developers with the high-signal, verifiable academic foundation required to build Large Language Models that are genuinely competent, culturally grounded, and intellectually fluent in African realities.

Live Corpus Statistics & Domain Breakdown

Academic Domain	Paper Count	Share	Key Focus Areas
Social Sciences & Economics	~205,772 papers	24.4%	African trade, monetary policy, education, governance, economic development
Medicine & Public Health	~198,240 papers	23.5%	Tropical epidemiology, clinical health, malaria, maternal care, infectious diseases
Agriculture & Food Security	~152,027 papers	18.0%	Crop protection, agronomy, soil fertility, CGSpace manuals, smallholder farming
Engineering & Earth Sciences	~125,140 papers	14.8%	Mining, hydrology, renewable energy, materials, geology, environmental engineering
Computer Science & AI	~80,020 papers	9.5%	African NLP, speech translation models, machine learning, computer vision, robotics
Humanities & Literature	~48,471 papers	5.7%	African linguistics, indigenous languages (Swahili, Yoruba, Hausa, Twi, Zulu), history, literature
Physical Sciences & Mathematics	~27,305 papers	3.2%	Astronomy (MeerKAT/SKA), pure & applied mathematics, theoretical physics, chemistry
Multidisciplinary / General Science	~5,237 papers	0.6%	General science, university research bulletins, multidisciplinary studies
Law & Governance	~2,595 papers	0.3%	African customary law, international human rights, constitutional jurisprudence
TOTAL (LIVE AUDIT)	844,807 papers	100%	10.15 BILLION TOKENS across 879 Parquet shards

Geographic Distribution: All 54 African Sovereign Nations

The corpus comprehensively indexes scholarly research originating from and focused on all 54 African sovereign nations. Rather than a partial sampling, the distribution covers every African member state:

Rank	African Country	Full-Text Papers	Share of Corpus (%)
1	Nigeria	184,891	21.89%
2	South Africa	155,133	18.36%
3	Egypt	83,837	9.92%
4	Algeria	48,925	5.79%
5	Kenya	47,908	5.67%
6	Morocco	47,881	5.67%
7	Ethiopia	41,574	4.92%
8	Ghana	27,101	3.21%
9	Tunisia	25,799	3.05%
10	Uganda	22,211	2.63%
11	Tanzania	17,476	2.07%
12	Cameroon	17,332	2.05%
13	Libya	13,532	1.60%
14	Côte d'Ivoire	10,780	1.28%
15	Sudan	10,754	1.27%
16	Senegal	9,412	1.11%
17	DR Congo	8,755	1.04%
18	Zambia	7,965	0.94%
19	Zimbabwe	7,563	0.90%
20	Burkina Faso	7,140	0.85%
21	Benin	6,668	0.79%
22	Rwanda	5,350	0.63%
23	Mozambique	5,300	0.63%
24	Malawi	4,751	0.56%
25	Botswana	3,322	0.39%
26	Madagascar	2,655	0.31%
27	Namibia	2,615	0.31%
28	Mauritius	1,083	0.13%
29	Mali	863	0.10%
30	Togo	688	0.08%
31	Gambia	439	0.05%
32	Niger	372	0.04%
33	Angola	324	0.04%
34	Burundi	294	0.03%
35	Somalia	174	0.02%
36	Guinea	98	0.01%
37	Sierra Leone	91	0.01%
38	Chad	88	0.01%
39	South Sudan	87	0.01%
40	Eswatini	78	0.01%
41	Eritrea	61	0.01%
42	Lesotho	56	0.01%
43	Liberia	56	0.01%
44	Congo	54	0.01%
45	Mauritania	44	0.01%
46	Gabon	38	0.00%
47	Djibouti	37	0.00%
48	Guinea-Bissau	26	0.00%
49	Central African Republic	25	0.00%
50	Cabo Verde	19	0.00%
51	Comoros	14	0.00%
52	Equatorial Guinea	12	0.00%
53	Seychelles	12	0.00%
54	Sao Tome and Principe	10	0.00%

Corpus Highlights & Scope

• DOAJ African Journals: Complete continental catalog of peer-reviewed open-access journals published directly by African universities and research institutes.
• African arXiv STEM & AI: Complete collection of African-authored or African-focused preprints in Machine Learning, NLP (cs.CL), Mathematics, and Physics.
• CGIAR CGSpace Agricultural Knowledge: Comprehensive crop disease guides, agronomy, soil fertility, and farming systems from IITA, ILRI, AfricaRice, and ICRAF.
• Global African Scholar Open Access: Curated open-access papers from African scientists publishing worldwide in Nature, PLOS, Springer, and international repositories.

How to Load in Python

from datasets import load_dataset

# 1. Load the entire dataset
dataset = load_dataset("AfriSpeech/africa-papers", split="train")

# 2. Or filter to a specific domain / source
agri = dataset.filter(lambda r: r["domain"] == "Agriculture & Food Security")
arxiv = dataset.filter(lambda r: r["source"] == "arXiv")

Dataset Fields

Column	Description
article_id	Unique paper identifier
title	Paper title
text	Full paper as clean Markdown (title, metadata line, abstract, body)
source	Collection: DOAJ, arXiv, CGSpace, or OpenAlex
domain	Standardized academic domain
country	African country of publication or author affiliation
journal	Journal or repository name
authors	Author names
year	Publication year
doi	Digital Object Identifier (when available)
url	Landing page or PDF URL
token_count	Token count of text (tiktoken cl100k_base)