pretty_name: HALvest
configs:
- config_name: bg
data_files: bg/*.gz
- config_name: br
data_files: br/*.gz
- config_name: ca
data_files: ca/*.gz
- config_name: cs
data_files: cs/*.gz
- config_name: da
data_files: da/*.gz
- config_name: de
data_files: de/*.gz
- config_name: el
data_files: el/*.gz
- config_name: en
data_files: en/*.gz
- config_name: eo
data_files: eo/*.gz
- config_name: es
data_files: es/*.gz
- config_name: et
data_files: et/*.gz
- config_name: eu
data_files: eu/*.gz
- config_name: fa
data_files: fa/*.gz
- config_name: fi
data_files: fi/*.gz
- config_name: fr
data_files: fr/*.gz
- config_name: gl
data_files: gl/*.gz
- config_name: he
data_files: he/*.gz
- config_name: hr
data_files: hr/*.gz
- config_name: hu
data_files: hu/*.gz
- config_name: hy
data_files: hy/*.gz
- config_name: id
data_files: id/*.gz
- config_name: it
data_files: it/*.gz
- config_name: ko
data_files: ko/*.gz
- config_name: 'no'
data_files: no/*.gz
- config_name: pl
data_files: pl/*.gz
- config_name: pt
data_files: pt/*.gz
- config_name: ro
data_files: ro/*.gz
- config_name: ru
data_files: ru/*.gz
- config_name: sk
data_files: sk/*.gz
- config_name: sl
data_files: sl/*.gz
- config_name: sv
data_files: sv/*.gz
- config_name: sw
data_files: sw/*.gz
- config_name: th
data_files: th/*.gz
- config_name: tr
data_files: tr/*.gz
language:
- bg
- br
- ca
- cs
- da
- de
- el
- en
- eo
- es
- et
- eu
- fa
- fi
- fr
- gl
- he
- hr
- hu
- hy
- id
- it
- ko
- 'no'
- pl
- pt
- ro
- ru
- sk
- sl
- sv
- sw
- th
- tr
size_categories:
- n<1K
- 1K<n<10K
- 10K<n<100K
- 100K<n<1M
task_categories:
- text-generation
- fill-mask
task_ids:
- language-modeling
- masked-language-modeling
tags:
- academia
- research
annotations_creators:
- no-annotation
multilinguality:
- multilingual
source_datasets:
- HALvest-R
HALvest
Open Scientific Papers Harvested from HAL
Dataset Description
- Repository: GitHub
Dataset Summary
overview:
This dataset is comprised of fulltext from open papers found on Hyper Articles en Ligne (HAL). Our dump is mostly english/french but gather papers written in 34 languages across 13 domains.
You can download the dataset using Hugging Face datasets:
from datasets import load_dataset
ds = load_dataset("Madjakul/HALvest", "en")
Details
Building the dataset is a four steps process: data fetching from HAL, data merging, data enriching and data filtering.
- We first request HAL's API in order to gather open research papers and parse it -- effectively sorting papers by language. Then, we download the PDFs of the fetched data.
- Using GROBID, we convert each PDF to an
xml-tei
format in order to have structured data. We convert eachxml-tei
file to atxt
format before concatenating it with the paper's. - We compute some statistics about each document.
- We filter the data based of off simple ratios to expurge badly encoded documents.
Languages
ISO-639 | Language | # Documents | # mT5 Tokens |
---|---|---|---|
en | English | 442,892 | 7,606,895,258 |
fr | French | 193,437 | 8,728,722,255 |
es | Spanish | 2,930 | 68,076,878 |
it | Italian | 1,172 | 48,747,986 |
pt | Portuguese | 934 | 32,918,832 |
de | German | 646 | 11,699,417 |
ru | Russian | 245 | 5,763,532 |
eu | Basque | 112 | 2,297,460 |
pl | Polish | 43 | 987,878 |
el | Greek | 42 | 1,680,696 |
ro | Romanian | 39 | 1,298,901 |
ca | Catalan | 28 | 975,078 |
da | Danish | 26 | 961,895 |
br | Breton | 24 | 998,088 |
ko | Korean | 17 | 226,268 |
tr | Turkish | 17 | 149,718 |
hu | Hungarian | 14 | 577,568 |
eo | Esperanto | 14 | 105,286 |
fa | Persian | 10 | 190,929 |
hy | Armenian | 10 | 127,988 |
cs | Czech | 9 | 712,263 |
bg | Bulgarian | 8 | 180,146 |
id | Indonesian | 9 | 53,075 |
he | Hebrew | 8 | 61,283 |
hr | Croatian | 8 | 40,621 |
et | Estonian | 7 | 20,405 |
sv | Swedish | 6 | 270,642 |
no | Norwegian | 6 | 62,767 |
fi | Finnish | 3 | 17,583 |
sw | Swahili | 2 | 73,921 |
gl | Galician | 2 | 29,688 |
th | Thai | 1 | 70,909 |
sl | Slovenian | 1 | 22,844 |
sk | Slovak | 1 | 12,997 |
Domains
Domain | Code | # Documents | # mT5 Tokens |
---|---|---|---|
Humanities and Social Sciences | shs | 152,818 | 5,487,738,344 |
Computer Science | info | 143,229 | 2,436,890,715 |
Life Sciences | sdv | 111,038 | 3,008,633,879 |
Engineering Sciences | spi | 99,393 | 2,155,602,249 |
Physics | phys | 63,557 | 1,435,905,328 |
Mathematics | math | 54,393 | 1,359,277,656 |
Chemical Science | chim | 38,500 | 857,617,219 |
Environmental Science | sde | 30,827 | 566,560,266 |
Sciences of the Universe | sdu | 22,917 | 654,909,131 |
Statistics | stat | 20,571 | 1,449,842,318 |
Cognitive science | scco | 11,584 | 222,832,732 |
Quantitative Finance | qfin | 3,290 | 64,970,285 |
Nonlinear Sciences | nlin | 1,908 | 29,296,684 |
You can browse through every domains and sub-domains here: https://hal.science/browse/domain.
Considerations for Using the Data
The corpus is extracted from the HAL's open archive which distributes scientific publications following open access principles. The corpus is made up of both creative commons licensed and copyrighted documents (distribution authorized on HAL by the publisher). This must be considered prior to using this dataset for any purpose, other than training deep learning models, data mining etc. We do not own any of the text from which these data has been extracted.
Citation
@software{almanach_halvest_2024,
author = {Kulumba, Francis and Antoun, Wissam and Vimont, Guillaume and Romary, Laurent},
title = {HALvest: Open Scientific Papers Harvested from HAL.},
month = April,
year = 2024,
company = Almanach,
url = {https://github.com/Madjakul/HALvesting}
}
Dataset Copyright
The licence terms for HALvest strictly follows the one from HAL. Please refer to the below license when using this dataset.