Hugging Face's logo

Datasets:
dungarr12
/
jora_corpus1

Tasks:
Text Generation
Modalities:
Text
Formats:
text
Languages:
French
English
Arabic
Size:
1B - 10B
Libraries:
Datasets
License:
Dataset card Data Studio Files
xet
 Community
1
Dataset Viewer (First 5GB)
Auto-converted to Parquet Duplicate
text
stringlengths
0
100k
What type of organism is commonly used in preparation of foods such as cheese and yogurt?
mesophilic organisms
Mesophiles grow best in moderate temperature, typically between 25Β°C and 40Β°C (77Β°F and 104Β°F). Mesophiles are often found living in or on the bodies of humans or other animals. The optimal growth temperature of many pathogenic mesophiles is 37Β°C (98Β°F), the normal human body temperature. Mesophilic organisms have impo...
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Changes from a less-ordered state to a more-ordered state (such as a liquid to a solid) are always what?
exothermic
Summary Changes of state are examples of phase changes, or phase transitions. All phase changes are accompanied by changes in the energy of a system. Changes from a more-ordered state to a less-ordered state (such as a liquid to a gas) areendothermic. Changes from a less-ordered state to a more-ordered state (such as a...
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
When a meteoroid reaches earth, what is the remaining object called?
meteorite
Meteoroids are smaller than asteroids, ranging from the size of boulders to the size of sand grains. When meteoroids enter Earth’s atmosphere, they vaporize, creating a trail of glowing gas called a meteor. If any of the meteoroid reaches Earth, the remaining object is called a meteorite.
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Organisms categorized by what species descriptor demonstrate a version of allopatric speciation and have limited regions of overlap with one another, but where they overlap they interbreed successfully?.
ring species
Ring species Ring species demonstrate a version of allopatric speciation. Imagine populations of the species A. Over the geographic range of A there exist a number of subpopulations. These subpopulations (A1 to A5) and (Aa to Ae) have limited regions of overlap with one another but where they overlap they interbreed su...
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Zinc is more easily oxidized than iron because zinc has a lower reduction potential. since zinc has a lower reduction potential, it is a more what?
active metal
One way to keep iron from corroding is to keep it painted. The layer of paint prevents the water and oxygen necessary for rust formation from coming into contact with the iron. As long as the paint remains intact, the iron is protected from corrosion. Other strategies include alloying the iron with other metals. For ex...
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
What tells you how much of the food you should eat to get the nutrients listed on the label?
serving size
The information listed at the right of the label tells you what to look for. At the top of the label, look for the serving size. The serving size tells you how much of the food you should eat to get the nutrients listed on the label. A cup of food from the label pictured below is a serving. The calories in one serving ...
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Boron only occurs naturally in compounds with what element?
oxygen
Occurrence, Preparation, and Compounds of Boron and Silicon Boron constitutes less than 0.001% by weight of the earth’s crust. In nature, it only occurs in compounds with oxygen. Boron is widely distributed in volcanic regions as boric acid, B(OH)3, and in dry lake regions, including the desert areas of California, as ...
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Fertilization is the union of a sperm and egg, resulting in the formation of what?
a zygote
Fertilization is the union of a sperm and egg, resulting in the formation of a zygote.
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
The plants alternation between haploid and diploud generations allow it to do what?
reproduce asexually and sexually
All plants have a characteristic life cycle that includes alternation of generations . Plants alternate between haploid and diploid generations. Alternation of generations allows for both asexual and sexual reproduction. Asexual reproduction with spores produces haploid individuals called gametophytes . Sexual reproduc...
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Cutting down on the use of chemical fertilizers and preserving wetlands are ways to prevent what "unlivable" regions in bodies of water?
dead zones
Cutting down on the use of chemical fertilizers is one way to prevent dead zones in bodies of water. Preserving wetlands is also important. Wetlands are habitats such as swamps, marshes, and bogs where the ground is soggy or covered with water much of the year. Wetlands slow down and filter runoff before it reaches bod...
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
This sharing of electrons produces what is known as a covalent bond. covalent bonds are ~20 to 50 times stronger than what?
van der waals interactions
any other electron, they become a part of the molecule’s electron system.204 This sharing of electrons produces what is known as a covalent bond. Covalent bonds are ~20 to 50 times stronger than van der Waals interactions. What exactly does that mean? Basically, it takes 20 to 50 times more energy to break a covalent b...
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Interstitial carbides are produced by the reaction of most transition metals at high temperatures with what element?
carbon
temperatures with electropositive metals such as those of groups 1 and 2 and aluminum produces ionic carbides, which contain discrete metal cations and carbon anions. The identity of the anions depends on the size of the second element. For example, smaller elements such as beryllium and aluminum give methides such as ...
━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━
Fungus-like protist saprobes play what role in a food chain and are specialized to absorb nutrients from nonliving organic matter, such as dead organisms or their wastes?
decomposers
End of preview. Expand in Data Studio

JORA Corpus 1 (jinofy-corp/jora_corpus1)

JORA Corpus 1 is the cleaned and deduplicated pre-training corpus built by JINOVAE Labs, the AI division of JINOVAE (Africa), through its JINOFY data pipeline, to train the JoraMind family of language models. It brings together web text, encyclopedic text, code, mathematics, medicine and agriculture in 18 languages, with a deliberate focus on African languages (Amharic, Hausa, Swahili, Somali, Yoruba, Malagasy, Lingala, Comorian) alongside major international languages.

The data is not a raw mirror of its sources: every subset was downloaded, then cleaned, normalized and deduplicated by JINOVAE Labs (see Processing below).

The repository has no single license. Every subset keeps the license of its original source, listed per folder below. Anyone reusing a subset must follow that subset's license, including attribution, share-alike and non-commercial terms where they apply.

At a glance

Total size β‰ˆ 2.0 TB of cleaned UTF-8 text (lower bound: the six largest folders are only partly measured, see below)
Folders 73 top-level folders under corpus/
Languages French, English, Arabic, Spanish, Portuguese, Hindi, Indonesian, Malay, Turkish, Urdu, Amharic, Hausa, Swahili, Somali, Yoruba, Malagasy, Lingala, Comorian
Domains General web, Wikipedia, source code (12 languages), code Q&A, mathematics (reasoning traces), medicine, agriculture, books
Format Plain-text batches (<folder>_batch_NNNN.txt, about 0.4–0.55 GB each); corpus/medical_raw/ uses JSON Lines (.jsonl)

Processing

Each subset went through the JINOFY pipeline of JINOVAE Labs before upload:

  • Cleaning: removal of markup, boilerplate, encoding errors and malformed or too-short documents; Unicode normalization.
  • Deduplication: removal of duplicate documents within each subset.
  • Packaging: documents written as plain-text batches of about 0.5 GB.

The licenses below apply to the underlying texts; the processing does not change them.

Repository layout 

corpus/
β”œβ”€β”€ <subset>/
β”‚   β”œβ”€β”€ <subset>_batch_0000.txt        # cleaned text, documents separated by blank lines
β”‚   β”œβ”€β”€ <subset>_batch_0001.txt
β”‚   β”œβ”€β”€ README.md                      # (some subsets) short provenance note
β”‚   └── _state.json / _checkpoint.*    # (some subsets) resumable-download state written by the collection scripts
β”œβ”€β”€ medical_raw/
β”‚   β”œβ”€β”€ medical_qa/   openfda_raw/   pubmed_qa/   wikipedia_fr/   (*.jsonl)
└── _wura_state.json                   # download state of the WURA subsets (amh, hausa, mlg, sw, yo)

Folder naming. Folder names follow the collection scripts, partly in French. francais = French, anglais = English, arabe = Arabic, espagnol = Spanish, portugais = Portuguese, indonesien = Indonesian, malais = Malay, turc = Turkish, ourdou = Urdu, amharique = Amharic, malgache = Malagasy, comorien = Comorian, mathematiques = mathematics, medecine = medicine. Several African languages also appear under ISO-style codes (am, amh, ha, sw, swh, yo, yor, mlg, ln, lin). These are separate downloads from different sources or runs, not aliases of one another.

State files. _state.json, _checkpoint.json, _checkpoint.txt and _c4_state.json record download progress, for example the last processed source file or the list of completed sources. They are not training data.

Subsets, sources and licenses

Legend. The License column gives the license declared by the original source. β€” means that no license is recorded for that subset: it was collected or assembled by JINOVAE Labs, or its upstream source does not state a license.

1. General web text

Folder Language Size Source License
web_fineweb English β‰₯ 96.2 GB HuggingFaceFW/fineweb ODC-By 1.0
web_fineweb_edu English β‰₯ 92.0 GB HuggingFaceFW/fineweb-edu ODC-By 1.0
web_falcon English β‰₯ 98.8 GB tiiuae/falcon-refinedweb ODC-By 1.0

2. C4 / mC4 (Common Crawl, per language)

All *_c4 and *_mc4 folders come from the cleaned Common Crawl corpora C4 / multilingual C4 published by AllenAI (allenai/c4). License: ODC-By 1.0. Use is also subject to the Common Crawl terms of use.

Folder Language Size
francais_c4 French β‰ˆ 98 GB (β‰₯ 184 batches)
anglais_c4 English β‰₯ 96.2 GB
hindi_c4 Hindi 96.5 GB
arabe_c4 Arabic 78.7 GB
turc_c4 Turkish 36.6 GB
indonesien_c4 Indonesian 31.2 GB
espagnol_c4 Spanish 28.1 GB
ourdou_c4 Urdu 21.3 GB
portugais_c4 Portuguese 21.3 GB
malais_c4 Malay 20.7 GB
sw_mc4 Swahili 12.7 GB
sw_c4 Swahili 4.8 GB
swahili_c4 Swahili 4.3 GB
amharique_c4 Amharic 2.6 GB
am_c4 Amharic 2.5 GB
somali_c4 Somali 2.3 GB
am_mc4 Amharic 2.2 GB
ha_c4 Hausa 2.0 GB
malgache_c4 Malagasy 1.3 GB
ha_mc4 Hausa 1.3 GB
yo_mc4 Yoruba 0.23 GB

3. Wikipedia

All *_wikipedia folders come from Wikipedia dumps (wikimedia/wikipedia). License: CC BY-SA 3.0 and GFDL (attribution and share-alike required). The same applies to medical_raw/wikipedia_fr (section 7).

Folder Language Size
anglais_wikipedia English 22.8 GB
ourdou_wikipedia Urdu 18.8 GB
espagnol_wikipedia Spanish 16.4 GB
francais_wikipedia French 15.7 GB
malais_wikipedia Malay 14.4 GB
portugais_wikipedia Portuguese 11.8 GB
arabe_wikipedia Arabic 9.1 GB
hindi_wikipedia Hindi 2.9 GB
indonesien_wikipedia Indonesian 1.1 GB
turc_wikipedia Turkish 1.0 GB
ha_wikipedia Hausa 0.30 GB
sw_wikipedia Swahili 0.26 GB
swahili_wikipedia Swahili 0.13 GB
malgache_wikipedia Malagasy 0.13 GB
am_wikipedia Amharic 0.056 GB
yo_wikipedia Yoruba 0.039 GB
amharique_wikipedia Amharic 0.039 GB
ln_wikipedia Lingala 0.005 GB

4. African-language corpora

Folder Language Size Source License
sw Swahili 11.5 GB castorini/wura (passages-v1.0/train/sw.txt) Apache-2.0
hausa Hausa 2.6 GB castorini/wura Apache-2.0
amh Amharic β‰ˆ 1.1 GB castorini/wura Apache-2.0
yo Yoruba 0.62 GB castorini/wura Apache-2.0
mlg Malagasy 0.47 GB castorini/wura Apache-2.0
swh_scrap_clean Swahili 1.7 GB JINOVAE Labs web collection (East African news and government sites), cleaned β€”
amh_scrap_clean Amharic 1.2 GB JINOVAE Labs web collection, cleaned β€”
lin_scrap Lingala 0.15 GB JINOVAE Labs web collection β€”
mlg_scrap_clean Malagasy 0.074 GB JINOVAE Labs web collection, cleaned β€”
haus_scrap_clean Hausa 0.064 GB JINOVAE Labs web collection, cleaned β€”
yor_scrap_clean Yoruba 0.060 GB JINOVAE Labs web collection, cleaned β€”
comorien_manuel Comorian 0.3 MB JINOVAE Labs, manually assembled β€”

The Hausa, Malagasy, Swahili, Yoruba and Amharic C4/mC4 and Wikipedia subsets are listed in sections 2 and 3.

5. Source code and code Q&A

Folder Content Size Source License
code_python Python 52.9 GB not recorded β€”
code_java Java 50.8 GB not recorded β€”
code_javascript JavaScript 50.2 GB not recorded β€”
code_go Go 30.8 GB not recorded β€”
code_rust Rust 30.5 GB not recorded β€”
code_cpp_c C / C++ 30.4 GB not recorded β€”
code_bash Shell 26.5 GB not recorded β€”
code_php PHP 23.8 GB not recorded β€”
code_sql SQL 23.1 GB not recorded β€”
code_python_codeparrot Python 22.6 GB codeparrot/codeparrot-clean (GitHub code) β€”
code_html_css HTML / CSS 20.2 GB not recorded β€”
code_github_top Mixed, popular GitHub repositories 9.1 GB not recorded β€”
qa_glaive Code assistant Q&A 2.2 GB Glaive AI datasets (glaive-code-assistant, glaive-function-calling-v2) Apache-2.0
qa_code_misc Mixed code Q&A 0.32 GB not recorded β€”

Source files in the code subsets come from public repositories. Each file remains under the license of the repository it was taken from.

6. Mathematics β€” mathematiques (79.8 GB, 150 batches, β‰ˆ 6.0 M documents)

Per its _state.json, this folder concatenates the following sources:

Source Content License
open-r1/OpenR1-Math-220k DeepSeek-R1 reasoning traces Apache-2.0
TIGER-Lab/MathInstruct 262k multi-source instructions MIT
meta-math/MetaMathQA 395k augmented reasoning MIT
AI-MO/NuminaMath-CoT 860k chain-of-thought solutions Apache-2.0
AI-MO/NuminaMath-1.5 Olympiad and competition problems Apache-2.0
nvidia/OpenMathInstruct-2 14M problem/solution pairs CC BY 4.0

7. Medicine β€” medecine (β‰ˆ 80 GB) and medical_raw (β‰ˆ 3 GB)

Per its _state.json, medecine concatenates the following sources:

Source License
PubMedQA β€” labeled (gold), unlabeled (silver) and artificial (bronze) MIT
MedMCQA β€” AIIMS/NEET exams (train and validation) Apache-2.0
EPFL Clinical Guidelines Custom ("other"; see the dataset card)
MedQuAD β€” NIH / MedlinePlus Q&A CC BY 4.0
PMC-Patients β€” 250k clinical case reports CC BY-NC-SA 4.0 (non-commercial)
Common Pile β€” peS2o β€” openly licensed academic papers (medical) Per-article open licenses
MedAlpaca Medical Meadow β€” WikiDoc, WikiDoc Patient Information Creative Commons (variant not specified on the card)
MedAlpaca Medical Meadow β€” MedQA, PubMed Causal, Health Advice, Medical Flashcards, CORD-19 β€”
USMLE self-assessment questions β€”
Lavita ChatDoctor β€” HealthCareMagic β€”
Medical Questions Pairs β€”
medical_raw/ subfolder Content Size License
openfda_raw U.S. FDA open data (openFDA) 0.15 GB Public domain / CC0 (openFDA terms)
wikipedia_fr Medical articles from French Wikipedia β‰ˆ 1.9 GB CC BY-SA 3.0 / GFDL
pubmed_qa PubMedQA sample 5.6 kB MIT
medical_qa Mixed medical Q&A 0.95 GB β€”

8. Agriculture and books

Folder Content Size Source License
agriculture Agriculture and food-security text, with African relevance (tropical crops, food security) 79.2 GB JINOVAE Labs collection, cleaned β€”
pile_books Books β‰₯ 94.9 GB The Pile (EleutherAI), books component; exact sub-component not recorded β€”

Licensing summary

License Subsets
ODC-By 1.0 all *_c4 / *_mc4, web_fineweb, web_fineweb_edu, web_falcon
CC BY-SA 3.0 + GFDL all *_wikipedia, medical_raw/wikipedia_fr
Apache-2.0 WURA (amh, hausa, mlg, sw, yo), qa_glaive; parts of mathematiques and medecine
MIT parts of mathematiques (MathInstruct, MetaMathQA) and medecine (PubMedQA); medical_raw/pubmed_qa
CC BY 4.0 parts of mathematiques (OpenMathInstruct-2) and medecine (MedQuAD)
CC BY-NC-SA 4.0 part of medecine (PMC-Patients) β€” non-commercial use only
Public domain / CC0 medical_raw/openfda_raw
No license recorded (β€”) JINOVAE Labs collections (*_scrap*, lin_scrap, comorien_manuel, agriculture, medical_raw/medical_qa) and subsets without a recorded or declared license (code folders, qa_code_misc, pile_books, several medical components)

Obligations to keep in mind.

  • ODC-By and CC BY require attribution to the original source.
  • CC BY-SA and GFDL (Wikipedia) require attribution and share-alike for redistributed text.
  • CC BY-NC-SA (PMC-Patients) forbids commercial use of that component.
  • Texts gathered by JINOVAE Labs from websites remain the property of their original publishers.

Known residual issues

Deduplication was applied within each subset. The following residual issues were found by inspecting the repository's file listings (content hashes) and early training runs:

  • Duplicated files. Some batches share identical content hashes, for example am_mc4 batches 0002 and 0003 (= 0000 and 0001), medical_raw/medical_qa (the same content repeats every 8 batches) and medical_raw/wikipedia_fr. A few ourdou_wikipedia and code_bash entries are also repeated. These identical batch files should be dropped.
  • Overlapping subsets. The same language can appear in several C4/mC4 folders (am_c4 / amharique_c4 / am_mc4, sw_c4 / swahili_c4 / sw_mc4, …) and in WURA, which is itself partly built from mC4. Expect cross-subset duplication.
  • Repetitive or memorizable text. Training on the French subsets showed stretches with near-zero loss (perplexity β‰ˆ 1.7 over millions of tokens). This is typical of templated or duplicated web pages and calls for an additional near-duplicate pass.
  • Small or near-empty folders. medical_raw/pubmed_qa (5.6 kB), comorien_manuel (0.3 MB) and ln_wikipedia (5 MB).

Intended use

Pre-training and continued pre-training of the JoraMind language models by JINOVAE Labs. The corpus is cleaned and deduplicated within each subset; it is not deduplicated across subsets and not balanced by language. JINOVAE Labs' training pipeline applies its own sampling, upsampling of African languages, and filtering.

Citation

If you use a subset, please cite its original source as listed above. For the compilation:

@misc{jora_corpus1,
  title  = {JORA Corpus 1: a multilingual, Africa-focused pre-training corpus},
  author = {JINOVAE Labs},
  year   = {2026},
  url    = {https://huggingface.co/datasets/jinofy-corp/jora_corpus1}
}
Downloads last month
-
Number of rows:
36,972,483
Estimated number of rows:
7,755,817,839
Total file size:
2.06 TB