OpenLLM-France/Lucie-Training-Dataset

Lucie Training Dataset Card

The Lucie Training Dataset is a curated collection of text data in English, French, German, Spanish and Italian culled from a variety of sources including: web data, video subtitles, academic papers, digital books, newspapers, and magazines, some of which were processed by Optical Character Recognition (OCR). It also contains samples of diverse programming languages.

The Lucie Training Dataset was used to pretrain Lucie-7B, a foundation LLM with strong capabilities in French and English. Code for data preparation can be found in the training respository for Lucie-7B. Due to the licenses of a few subcorpora, the Lucie Training Dataset is released under a CC BY-NC-SA 4.0. A subset available for commercial use will be released soon.

We note that one subcorpus used for training could not be released with the Lucie Training Dataset due to copyright conflicts discovered after training had begun. This data came from the Persée platform. The full list of urls used to create the dataset can be recreated from the file persee_metadata_documents.csv, where the corresponding url is https://www.persee.fr/doc/{ID} for each ID in the column file_id. The file persee_metadata_collections.csv gives statistics on document, word and character counts for the data grouped by collection. In all, the corpus contains a total of 3.25 billion words and 5.75 billion tokens, making up around 0.25% of the raw corpus and 0.37% of the tokens seen during training.

Table of Contents:

• Dataset Description
  • Sample Metadata
  • Dataset Composition

    Web Newspaper Technical Book

    Legislative Texts Legislative Transcripts Wiki Math

    Forum Dialogue Multilingual Parallel Corpora Programming

  • Configurable Subsets and Versions
  • Details on Data Sources

    AmendementsParlement AmericanStories Claire (French and English) CroissantAligned DiscoursPublics Europarl and EuroparlAligned Eurovoc FineWebEdu GallicaMonographies

    GallicaPress Gutenberg HAL InterventionsParlement LEGI MathPile (Commercial) OpenData OpenEdition PeS2o (v2)

    Pile (Uncopyrighted) QuestionsEcritesParlement RedPajama (v2) Stac TheStack (v1.2) Theses Wikipedia, Wikisource, Wiktionary YouTube

• Example use in Python
• Load the dataset
• Iterate over a subset
• Load a specific version
• Citation
• Acknowledgements
• Contact

Dataset Description

This dataset is intended to provide extensive and diverse multilingual data for training Large Language Models (LLMs). Here are some of the principal features of the corpus:

• Data mix:
  • The dataset contains more French than English data -- it is in fact one of the biggest collections of French text data that has been preprocessed for LLM training -- with the aim of minimizing anglo-centric cultural biases.
  • German, Spanish and Italian are also represented in small amounts.
  • Code is included to boost the reasoning capabilities of LLMs.
• Data filtering and deduplication:
  • The dataset has been cleaned in an effort to remove very low-quality data.
  • Duplicate data samples have been removed to some extent, following best practices.
  • Web data has been filtered to minimize potentially toxic content and personally identifying information.
• Ethics:
  • Special care has been taken to respect copyright laws and individual privacy. All newspapers, monographies, magazines and legislative documents, as well as most books, are in the public domain (which depends on the author's date of death and the country of publication). Other data are published with permissive licenses (e.g., CC BY or CC BY-SA) or, in very rare cases, CC BY-NC-SA.
  • All web data in the dataset come from sites with robots.txt files that do not forbid crawling.

Sample Metadata

In addition to the text field, which provides the content of the sample, each training sample in the corpus contains the following metadata when available:

• language: the language of the text sample (note that this information is taken from the original data source and may be incorrect).
  Possible values:
  • the ISO 639-1 code for a given natural language ("en", "fr", "de", "es", or "it"),
  • the name of a programming language prefixed by "code:" ("code:python", "code:c++", …), or
  • a list of ISO 639-1 codes separated by commas for data containing parallel translations ("fr,en", "de,fr", "es,en", "it,en", or one of those pairs in the opposite order if the languages appear in the opposite order in the text).
• source: an identifier for the source(s) of the text sample (Wikipedia, RedPajama, Gutenberg, …). All sources are described in detail below.
• id: an identifier that is unique among documents from the same source.
• url (optional): the URL of the original text sample on the web, if available.
• title (optional): the title of the original text sample, if available.
• author (optional): the author of the original text sample, if available.
  Note:

  The author name is given in plain text, except in the case of Gutenberg books, where it is the JSON serialized object of the author metadata.
• date (optional): the publication date of the original text sample, if available.
  Note:

  The text format of the date depends on the source.
• quality_signals (optional): a list of quality signals for the text sample in JSON format (which could be used for further filtering or sample weighting).
  Note:

  It can include indicators computed by `fasttext` and `CCNet`, statistics about occurrences of characters, words, special characters, etc.
• extra (optional): extra information about the text sample, in JSON format. This can include metadata about the source subset, the rights, etc.

The list of metadata available for each source is provided (without the text field) in metadata_examples.json.

Dataset Composition

The following figure shows the distribution of the dataset by language (colors) and category (hatch patterns).

The following table provides an overview of the dataset composition, broken down by source and language. Sources are grouped by category. The table provides the numbers of documents, words, tokens, and characters for each subset. All numbers in this table are available in the CSV file dataset_composition.csv. Token counts are computed using the tokenizer for Lucie-7B.

Subset	Language	M docs	B words	B tokens	B chars
TOTAL		2186.562	1356.021	2314.862	8842.200
	French (fr)	653.812	583.687	928.618	3619.672	composition details
	English (en)	554.289	412.202	611.894	2553.541	composition details
	code	125.769	51.306	228.954	630.749	composition details
	German (de)	165.915	105.609	206.610	764.779	composition details
	Spanish (es)	171.651	123.857	200.825	759.457	composition details
	Italian (it)	99.440	62.051	112.031	404.454	composition details
	fr-en	410.032	17.016	25.494	107.658	composition details
	it-en	1.901	0.100	0.151	0.638
	es-en	1.961	0.103	0.143	0.631
	de-fr	1.792	0.0908	0.141	0.621
Category: Web
RedPajama	French (fr)	640.770	477.758	741.023	2974.596	composition details
	German (de)	162.779	103.078	201.371	747.631	composition details
	Spanish (es)	169.447	121.751	197.125	746.984	composition details
	Italian (it)	97.324	60.194	108.416	393.012	composition details
FineWebEdu	English (en)	421.209	327.453	467.837	2018.215	composition details
Category: Newspaper
GallicaPress	French (fr)	3.205	67.496	121.606	408.882
AmericanStories	English (en)	59.420	8.902	14.313	50.844	composition details
Category: Technical
PeS2o	English (en)	38.972	42.296	65.365	268.963
HAL	French (fr)	0.349	9.356	16.224	58.308
Theses	French (fr)	0.102	7.547	14.060	47.758
Pile (USPTO_Backgrounds)	English (en)	5.139	3.492	5.105	22.309
OpenEdition	French (fr)	0.939	2.225	3.604	14.459
Pile (PhilPapers)	English (en)	0.0308	0.363	0.618	2.304
Pile (NIH_ExPorter)	English (en)	0.914	0.288	0.431	1.979
Category: Book
GallicaMonographies	French (fr)	0.278	15.106	25.169	90.456
Gutenberg	English (en)	0.0563	3.544	5.516	20.579
	French (fr)	0.00345	0.227	0.383	1.392
	German (de)	0.00188	0.0987	0.193	0.654
	Italian (it)	0.000958	0.0657	0.129	0.414
	Spanish (es)	0.000735	0.0512	0.0920	0.303
Category: Legislative Texts
Pile (FreeLaw)	English (en)	3.415	8.204	14.011	52.580
Eurovoc	English (en)	0.272	1.523	2.571	9.468
	Italian (it)	0.245	0.731	1.527	4.867
	German (de)	0.247	0.678	1.497	4.915
	Spanish (es)	0.246	0.757	1.411	4.684
OpenData	French (fr)	1.169	0.755	1.209	4.638
QuestionsEcritesParlement	French (fr)	0.189	0.108	0.156	0.705
LEGI	French (fr)	0.621	0.0878	0.145	0.563
AmendementsParlement	French (fr)	0.673	0.0452	0.0738	0.274
Category: Legislative Transcripts
Europarl	German (de)	0.0102	0.0451	0.0734	0.327
	Spanish (es)	0.0103	0.0524	0.0733	0.325
	French (fr)	0.0103	0.0528	0.0717	0.339
	English (en)	0.0111	0.0563	0.0690	0.339
DiscoursPublics	French (fr)	0.110	0.163	0.238	1.025
InterventionsParlement	French (fr)	1.832	0.104	0.157	0.654
Category: Wiki
Wikipedia	English (en)	6.893	4.708	7.898	26.616
	German (de)	2.877	1.709	3.476	11.252
	French (fr)	2.648	1.726	2.940	9.879
	Spanish (es)	1.947	1.245	2.124	7.161
	Italian (it)	1.870	1.060	1.959	6.161
wikisource	French (fr)	0.186	0.523	0.795	3.080
wiktionary	French (fr)	0.650	0.0531	0.117	0.347
Category: Math
MathPile	English (en)	0.737	3.408	9.637	27.290
Pile (DM_Mathematics)	English (en)	0.992	1.746	4.928	8.127
Category: Forum
Pile (StackExchange)	English (en)	15.269	4.534	10.275	33.609
Pile (Ubuntu_IRC)	English (en)	0.0104	0.867	2.159	5.610
Category: Dialogue
Claire	English (en)	0.949	0.818	1.161	4.709	composition details
	French (fr)	0.0393	0.210	0.311	1.314	composition details
YouTube	French (fr)	0.0375	0.145	0.336	1.003
STAC	English (en)	0.0000450	0.0000529	0.000121	0.000327
Category: Multilingual Parallel Corpora
CroissantAligned	fr-en	408.029	16.911	25.351	107.003
EuroparlAligned	it-en	1.901	0.100	0.151	0.638
	fr-en	2.003	0.105	0.143	0.655
	es-en	1.961	0.103	0.143	0.631
	de-fr	1.792	0.0908	0.141	0.621
Category: Programming
TheStack	JAVASCRIPT	21.109	8.526	58.609	141.647
	JAVA	20.152	7.421	27.680	89.297
	C	8.626	5.916	24.092	57.428
	PHP	15.905	4.865	22.883	66.844
	PYTHON	12.962	5.434	21.683	64.304
	C++	6.378	4.584	18.835	50.892
	C#	10.839	3.574	13.381	46.286
	GO	4.730	2.735	10.262	25.738
	TYPESCRIPT	10.637	2.617	9.836	28.815
	RUST	1.387	0.872	3.241	9.529
	RUBY	3.405	0.646	2.392	7.139
	SWIFT	1.756	0.553	1.876	6.134
	KOTLIN	2.243	0.454	1.758	5.769
	SCALA	1.362	0.457	1.587	4.862
	TEX	0.398	0.394	1.507	3.805
	LUA	0.559	0.318	1.367	3.279
	DART	0.933	0.308	1.242	3.864
	PERL	0.392	0.297	1.149	2.634
	MATHEMATICA	0.0269	0.120	1.117	1.720
	ASSEMBLY	0.248	0.209	0.867	1.575
	HASKELL	0.545	0.307	0.807	2.364
	FORTRAN	0.165	0.192	0.780	1.843
	JULIA	0.299	0.152	0.660	1.539
	OCAML	0.160	0.130	0.430	1.107
	ERLANG	0.0994	0.0657	0.260	0.726
	ELIXIR	0.282	0.0731	0.258	0.737
	CLOJURE	0.126	0.0448	0.179	0.492
	R	0.0392	0.0278	0.158	0.305
	MATLAB	0.000967	0.00865	0.0427	0.0372
	RACKET	0.00420	0.00479	0.0153	0.0378

Configurable Subsets and Versions

As the Lucie Training Dataset is a collection of multilingual corpora from different sources, it can be divided into subsets based on the source and language of its constituent corpora.
The list of possible configurations is available in the YAML header of this README file. Each configuration corresponds to a pathname pattern in the data subdirectory.

The dataset is also available in the following versions:

• v1.1 / main (default): The data used for the first (main) pretraining phase of Lucie-7B, which contains approximately 2.3T tokens. The statistics above apply to this version.
• v1.2: An improved version of the main dataset, where
  • GallicaMonographies and GallicaPress have been filtered aggressively to remove documents with low OCR quality. After filtering, GallicaMonographies contains around 220,000 documents and 20.131 billion tokens. For GallicaPress, we first selected a subset of the original corpus that contained only html documents (as opposed to documents in .txt format). This subset contained 1,747,600 documents and 74 billion tokens. After filtering, this subset contains roughly 989,100 documents and 45.7 billion tokens.
  • The Ubuntu_IRC and PhilPapers subsets of the Pile have been refined by fixing encoding issues and removing documents in languages other than English, French, Spanish, German and Italian. After filtering, Ubuntu_IRC contains about 9,000 documents and 1.745 billion tokens. PhilPapers contains around 28,000 million documents and 502 million tokens.
• v1.2-recent-web : The data used for the second pretraining phase (context extension) of Lucie-7B. This version is identical to v1.2 with the exception that older snapshots of web data (before 2023 for RedPajama and before 2024 for FineWebEdu) have been excluded. All data from v1.1 that were not filtered out remain unchanged in v1.2 and v1.2-recent-web.

Except from v1.1, which is a git tag, all versions are git branches in the dataset repository (e.g. v1.2).

The Example use in Python section contains example Python code for loading and iterating over the dataset with different configurations, including source, language and version.

Details on Data Sources

AmendementsParlement

• Source: Corpus contributed by OpenLLM partners.
• Extracted from: Regards citoyens. License: CC BY-SA.
• Description: A collection of proposed amendments by the French parliament. Documents contain the text of the proposed amendment, the name of the associated law as well as information on who voted on the amendment and what was decided.

AmericanStories

• Source: dell-research-harvard/AmericanStories. License: CC BY 4.0.
• Extracted from: Chronicling America. License: Open.
• Description: "The American Stories dataset is a collection of full article texts extracted from historical U.S. newspaper images. It includes nearly 20 million scans from the public domain Chronicling America collection maintained by the Library of Congress. The dataset is designed to address the challenges posed by complex layouts and low OCR quality in existing newspaper datasets" (from the dataset card). See the dataset composition details for statistics on documents by year. Dataset containing text retrieved through OCR.
• Pre-processing:
  • Filtering: To filter out documents with excessive OCR errors, the dataset was refined by discarding texts with a perplexity higher than 2310, measured using a CCNET model in English (see code details). The code to compute CCNET perplexity, parallelizing on parquet files, is available here.
• Citation: Melissa Dell, Jacob Carlson, Tom Bryan, Emily Silcock, Abhishek Arora, Zejiang Shen, Luca D'Amico-Wong, Quan Le, Pablo Querubin and Leander Heldring (2023). "American Stories: A Large-Scale Structured Text Dataset of Historical U.S. Newspapers," arxiv:2308.12477.

Claire (French and English)

• Sources:
  • French dataset: OpenLLM-France/Claire-Dialogue-French-0.1. License: CC BY-NC-SA 4.0.
  • English dataset: OpenLLM-France/Claire-Dialogue-English-0.1. License: CC BY-NC-SA 4.0.
• Extracted from: see the datacards for the French and English datasets.
• Description: The Claire datasets are composed of transcripts of spoken conversations -- including parliamentary proceedings, interviews, debates, meetings, and free conversations -- as well as some written conversations from theater plays and written chats. The dataset is designed to help downstream performance of models fine-tuned for tasks requiring the comprehension of spontaneous spoken conversation, such as meeting summarization. Each dialogue is split into speech turns, and each speech turn is labeled with the name of the speaker or a unique identifier. See the composition details for the French dataset and the English dataset for a high-level view of the distribution of different types of documents in each dataset.
• Citation: Julie Hunter, Jérôme Louradour, Virgile Rennard, Ismaïl Harrando, Guokan Shang, Jean-Pierre Lorré (2023). The Claire French Dialogue Dataset. arXiv:2311.16840.

CroissantAligned

• Source: croissantllm/croissant_dataset_no_web_data (subset: aligned_36b). License: not specified.
• Extracted from:
  • Translation pairs: OPUS (99.6% of the data in CroissantAligned). Pairs extracted from OPUS are labeled as "UnbabelFrEn".
  • Thesis abstracts: French thesis abstract pairs. License: ETALAB-Licence-Ouverte-v2.0.
  • Song lyrics: lacoccinelle.
• Description: CroissantAligned contains samples of parallel French/English (or English/French) data. Data extracted from OPUS takes the form of sentences pairs, where one sentence is in French and the other is in English. OPUS pairs were passed through a custom pipeline designed to select the highest quality translation examples. Selected pairs are labeled "UnbabelFrEn" in the CroissantAligned dataset. The thesis abstract subset contains thesis abstracts paired with translations written by the thesis authors. The song lyrics are translated by contributors to www.lacoccinelle.net. Parallel data are used to boost the multilingual capabilities of models trained on them (Faysse et al.,2024).
• Pre-processing:
  • Language separation and tagging: The original text field of the Croissant dataset contains a sentence or passage in French or English immediately followed by its translation without any indication of which passage is in which language. The first step was thus to split each text into separate, monolingual passages and tag each passage with the appropriate language code, identified automatically using the langid library (see code details). In the Lucie Training Dataset, the extra metadata field for CroissantAligned contains separate keys, text_fr for French and text_en for English, that stores the texts separately.
  • Random combination of texts prefixed by language: To create the text values, each monolingual text was repaired with its translation, but random separators and various methods of prefixing the text with the language (name or code) were added. This was done as a precaution to prevent models trained on this data from switching languages when generating text and can be seen as a very basic instruction to translate the source (first) text into the target (second) text (see code details).
• Citation: Manuel Faysse, Patrick Fernandes, Nuno M. Guerreiro, António Loison, Duarte M. Alves, Caio Corro, Nicolas Boizard, João Alves, Ricardo Rei, Pedro H. Martins, Antoni Bigata Casademunt, François Yvon, André F.T. Martins, Gautier Viaud, Céline Hudelot, Pierre Colombo (2024). "CroissantLLM: A Truly Bilingual French-English Language Model," arXiv:2402.00786.

DiscoursPublics

• Source: Corpus contributed by OpenLLM partners.
• Extracted from: Vie Publique. License: ETALAB-Licence-Ouverte-v2.0.
• Description: A collection of public speeches from the principal public actors in France including speeches from the French President starting from 1974 and from the Prime Minister and members of the government starting from 1980.
• Pre-processing:
  • Text cleaning: the mention of the source url and the number of views were removed from the text.

Europarl and EuroparlAligned

• Sources:
  • fr-en, es-en, it-en parallel data: Europarl v7. License: Open.
  • fr, en, de, es monolingual data and de-fr parallel data: Europarl v10. License: Open.
• Description: "The Europarl parallel corpus is extracted from the proceedings of the European Parliament. It includes versions in 21 European languages: Romanic (French, Italian, Spanish, Portuguese, Romanian), Germanic (English, Dutch, German, Danish, Swedish), Slavik (Bulgarian, Czech, Polish, Slovak, Slovene), Finni-Ugric (Finnish, Hungarian, Estonian), Baltic (Latvian, Lithuanian), and Greek. The goal of the extraction and processing was to generate sentence aligned text for statistical machine translation systems" (www.statmt.org).
• Pre-processing:
  • Random combination of aligned texts prefixed by language: The same process as used for the CroissantAligned dataset was applied to the EuroparlAligned dataset (see code details). In the Lucie Training Dataset, the extra field in the metadata for EuroparlAligned provides texts in the two languages under the sub-fields text_1 and text_2, and the corresponding language codes under lang_1 and lang_2.
• Citation: Philipp Koehn (2005). "Europarl: A Parallel Corpus for Statistical Machine Translation," MT Summit.

Eurovoc

• Source: EuropeanParliament/Eurovoc. License: EUPL 1.1.
• Extracted from: Cellar. License: CC BY-4.0.
• Description: A collection of mutlilingual documents from the data repository of the Publications Office of the European Union annotated with Eurovoc labels. The corpus contains legal, policy-related, historical and organizational information about the EU. Dataset containing text retrieved through OCR.
• Pre-processing:
  • Filtering: To filter out documents with excessive OCR errors, the dataset was refined by discarding texts with a perplexity higher than 1500, measured using a CCNET model on the target language (see code details). The code to compute CCNET perplexity, parallelizing on parquet files, is available here.
  • Text cleaning: Mentions of Credit Institutions Directives (CID) that appears in the raw texts such as (cid:146) were removed.
• Citations:
  • Ilias Chalkidis, Emmanouil Fergadiotis, Prodromos Malakasiotis, Nikolaos Aletras, and Ion Androutsopoulos (2019). "Extreme Multi-Label Legal Text Classification: A Case Study in EU Legislation," Proceedings of the Natural Legal Language Processing Workshop 2019, pages 78–87, Minneapolis, Minnesota. Association for Computational Linguistics.
  • Ilias Chalkidis, Manos Fergadiotis, Prodromos Malakasiotis and Ion Androutsopoulos (2019). "Large-Scale Multi-Label Text Classification on EU Legislation," Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics (ACL 2019), Florence, Italy, (short papers).
  • Andrei-Marius Avram, Vasile Pais, and Dan Ioan Tufis (2021). "PyEuroVoc: A Tool for Multilingual Legal Document Classification with EuroVoc Descriptors," Proceedings of the International Conference on Recent Advances in Natural Language Processing (RANLP 2021), pages 92–101, Held Online. INCOMA Ltd.
  • Zein Shaheen, Gerhard Wohlgenannt and Erwin Filtz (2020). "Large scale legal text classification using transformer models," arXiv:2010.12871.

FineWebEdu

• Source: HuggingFaceFW/fineweb-edu. License: ODC-BY.
• Extracted from: FineWeb. License: ODC-BY.
• Description: A 1.3 trillion token selection from FineWeb, which contains 15 trillion tokens of curated data from 96 Common Crawl dumps. Content in FineWebEdu has been selected by a custom designed classifier for its high-quality, educational content. Most recent crawl: 2024-10 (see composition details for information about the crawls included in this dataset.)
• Pre-processing:
  • Removing duplicate urls: urls were removed if their base domain overlapped with a dataset already in the Lucie Training Dataset (e.g., "philpapers.org") in order to increase diversity of content (see code details)
  • Filtering by robots.txt files: we collect robots.txt and remove all documents for which CCBot is disallowed or for which we failed to collect information as of July 2024 in an effort to select data free from opt-out evidence according to the 4th article of the copyright European directive (2019).
• Citation: Guilherme Penedo, Hynek Kydlíček, Loubna Ben allal, Anton Lozhkov, Margaret Mitchell, Colin Raffel, Leandro Von Werra, Thomas Wolf (2024). "The FineWeb Datasets: Decanting the Web for the Finest Text Data at Scale," arXiv:2406.17557.

GallicaMonographies

• Source: Corpus contributed by OpenLLM partners. A version is also published here: PleIAs/French-PD-Books. License: Public domain.
• Extracted from: Gallicagram.
• Description: A large collection of French monographies in the public domain made available through the French National Library (Gallica). Dataset containing text retrieved through OCR.
• Pre-processing:
  • Text cleaning for v1.1: To filter out documents with excessive OCR errors, the dataset was split into chunks and chunks were kept if the source language was detected as French by FastText with a confidence score of 0.65 or above, and the perplexity score, as measured using a CCNET model in French, was between 10 and 1000. The code to compute CCNET perplexity, parallelizing on parquet files, is available here.
  • Filtering for v1.2: Using OCR scores provided in the metadata of the source corpus, documents with an OCR score of less than 90 out of 100 were filtered out.

GallicaPress

• Source: Corpus contributed by OpenLLM partners. A version is also published here: PleIAs/French-PD-Newspapers. License: Public domain.
• Extracted from: Gallicagram.
• Description: A large collection of French newspapers and periodicals in the public domain made available through the French National Library (Gallica). Dataset containing text retrieved through OCR.
• Pre-processing:
  • Text cleaning for v1.1: To filter out documents with excessive OCR errors, the dataset was split into chunks and chunks were kept if the source language was detected as French by FastText with a confidence score of 0.65 or above, and the perplexity score, as measured using a CCNET model in French, was between 10 and 1000 (see code details). The code to compute CCNET perplexity, parallelizing on parquet files, is available here.
  • Filtering for v1.2: Using OCR scores provided in the metadata of the source corpus, documents with an OCR score of less than 90 out of 100 were filtered out.

Gutenberg

• Source: Corpus compiled by OpenLLM partners.
• Extracted from:
  • aleph.gutenberg.org via Project Gutenberg. License: Open.
  • pgcorpus. License: CC BY-4.0.
• Description: A collection of free eBooks, manually prepared by human annotators.
• Pre-processing:
  • Filtering: The dataset was filtered based on the author date of death, so that only texts from authors who died more than 70 years ago are included (80 years for French authors). See code details here. This filtering was done to ensure that the texts are in the public domain.
  • Text cleaning: Headers and footers containing information about Project Gutenberg were removed (see code details).

HAL

• Source: bigscience-data/roots_fr_hal_archives_ouvertes. License: Roots dataset.
• Extracted from: HAL (Open access).
• Description: A collection of scientific papers and manuscripts distributed through the open science platform HAL. Dataset containing text retrieved through OCR.
• Pre-processing:
  • Filtering: To filter out documents with excessive OCR errors, the dataset was refined by discarding texts with a perplexity higher than 930, measured using a CCNET model in French (see code details). The code to compute CCNET perplexity, parallelizing on parquet files, is available here.
• Citation: Hugo Laurençon, Lucile Saulnier, Thomas Wang, Christopher Akiki, Albert Villanova del Moral, Teven Le Scao, Leandro Von Werra, Chenghao Mou, Eduardo González Ponferrada, Huu Nguyen, Jörg Frohberg, Mario Šaško, Quentin Lhoest, Angelina McMillan-Major, Gerard Dupont, Stella Biderman, Anna Rogers, Loubna Ben allal, Francesco De Toni, Giada Pistilli, Olivier Nguyen, Somaieh Nikpoor, Maraim Masoud, Pierre Colombo, Javier de la Rosa, Paulo Villegas, Tristan Thrush, Shayne Longpre, Sebastian Nagel, Leon Weber, Manuel Muñoz, Jian Zhu, Daniel Van Strien, Zaid Alyafeai, Khalid Almubarak, Minh Chien Vu, Itziar Gonzalez-Dios, Aitor Soroa, Kyle Lo, Manan Dey, Pedro Ortiz Suarez, Aaron Gokaslan, Shamik Bose, David Adelani, Long Phan, Hieu Tran, Ian Yu, Suhas Pai, Jenny Chim, Violette Lepercq, Suzana Ilic, Margaret Mitchell, Sasha Alexandra Luccioni, Yacine Jernite (2022). "The BigScience ROOTS Corpus: A 1.6TB Composite Multilingual Dataset," Advances in Neural Information Processing Systems (NeurIPS), 35, 31809-31826.

InterventionsParlement

• Source: Corpus contributed by OpenLLM partners.
• Extracted from: Regards citoyens. License: CC BY-SA.
• Description: Transcripts of remarks made during French parlementary debates. Each text contains a continuous remark by a single speaker.

LEGI

• Source: Corpus contributed by OpenLLM partners. A version is also published here: Nicolas-BZRD/DILA_OPENDATA_FR_2023.
• Extracted from: OpenData (Data collection date: October, 2023).
• Description: "The French Government Open Data (DILA) Dataset is a collection of text data extracted from various sources provided by the French government, specifically the Direction de l'information légale et administrative (DILA). This dataset contains a wide range of legal, administrative, and legislative documents. The data has been organized into several categories for easy access and analysis" (from the dataset card).

MathPile (Commercial)

• Source: GAIR/MathPile_Commercial. License: CC BY-SA 4.0.
• Extracted from: MathPile. License: CC BY-SA-NC 4.0.
• Description: A preprocessed collection of documents focused on math, including Textbooks, arXiv, Wikipedia, ProofWiki, StackExchange, and web pages from Common Crawl. The content targets a range of levels, from kindergarten through postgraduate level. MathPile_Commercial was obtained by removing documents from MathPile that do not allow commercial use.
• Pre-processing:
  • Formatting: Converted the content of StackExchange questions and answers to match the {"text": value} format, using the following formula:

  text = sample["question"]["Body"] + "\n\n".join([answer["Body"] for answer in sample["answers"]])
  

• Citation: Zengzhi Wang, Rui Xia and Pengfei Liu (2023). "Generative AI for Math: Part I -- MathPile: A Billion-Token-Scale Pretraining Corpus for Math," arXiv:2312.17120.

OpenData

• Source: Nicolas-BZRD/DILA_OPENDATA_FR_2023 (balo, dole, inca, kali, and sarde subsets). License: ODC-BY.
• Extracted from: OpenData (Data collection date: October, 2023).
• Description: "The French Government Open Data (DILA) Dataset is a collection of text data extracted from various sources provided by the French government, specifically the Direction de l'information légale et administrative (DILA). This dataset contains a wide range of legal, administrative, and legislative documents. The data has been organized into several categories for easy access and analysis" (from the dataset card).

OpenEdition

• Source: Corpus contributed by OpenLLM partners.
• Extracted from: Open Edition. License: Open Edition Books.
• Description: A collection of scientific books, journal articles, blog entries and event descriptions.

PeS2o (v2)

• Source: allenai/peS2o version v2. License: ODC BY-v1.0.
• Extracted from: S2ORC (see aclanthology). License: ODC BY-v1.0.
• Description: A preprocessed collection of academic papers designed for pre-training of language models. PeS2o is composed of two subsets: one containing full papers and one containing only paper titles and abstracts. Dataset containing (some) text retrieved through OCR. Knowledge cutoff: 2023-01-03.
• Citation: Luca Soldaini and Kyle Lo (2023). "peS2o (Pretraining Efficiently on S2ORC) Dataset," Allen Institute for AI. GitHub.

Pile (Uncopyrighted)

• Source: monology/pile-uncopyrighted. License: Other.
• Extracted from: FreeLaw, StackExchange, USPTO Backgrounds, DM Mathematics, Ubuntu IRC, PhilPapers, NIH ExPorter from The Pile. License: MIT.
• Description (from the Datasheet):
  • FreeLaw: "The Free Law Project is US registered non-profit that provide access to millions of legal opinions and analytical tools for academic studies in the legal realm."
  • StackExchange: "The StackExchange dataset is a dump of anonymized user-contributed content on the Stack Exchange network, a popular collection of websites centered around user-contributed questions and answers."
  • USPTO Backgrounds: "The USPTO Backgrounds dataset is a set of background sections from patents granted by the United States Patent and Trademark Office, derived from its published bulk archives."
  • DM Mathematics: "The DeepMind Mathematics dataset consists of a collection of mathematical problems such as algebra, arithmetic, calculus, number theory, and probability, formatted as natural language prompts Saxton et al., 2019."
  • Ubuntu IRC: "The Ubuntu IRC dataset is derived from the publicly available chatlogs of all Ubunturelated channels on the Freenode IRC chat server."
  • PhilPapers: a dataset of open access philosophy publications from an international database maintained by the Center for Digital Philosophy at the University of Western Ontario.
  • NIH ExPORTER: "The NIH Grant abstracts provides a bulk-data repository for awarded applications through the ExPORTER4 service covering the fiscal years 1985-present."
• Pre-processing (v1.2 only):
  • Filtering of PhilPapers: Papers were removed if their language, detected using Stanza, was not classified as English, French, German, Spanish or Italian.
  • Filtering and text cleaning of Ubuntu IRC: Texts from some channels were excluded to avoid data from languages other than English, French, German, Spanish or Italian and certain encoding errors were fixed (see code details here).
• Citations:
  • Leo Gao, Stella Biderman, Sid Black, Laurence Golding, Travis Hoppe, Charles Foster, Jason Phang, Horace He, Anish Thite, Noa Nabeshima, Shawn Presser, Connor Leahy (2020). "The Pile: An 800GB Dataset of Diverse Text for Language Modeling," arXiv:2101.00027.
  • Stella Biderman, Kieran Bicheno, Leo Gao (2022). "Datasheet for the Pile," arXiv:2201.07311.

QuestionsEcritesParlement

• Source: Corpus contributed by OpenLLM partners.
• Extracted from: Regards citoyens. License: CC BY-SA.
• Description: Collection of long written questions, read during a session at the French National Assembly. Questions are asked by a member of the French parliament and addressed to a minister (who is given two months to respond).

RedPajama (v2)

• Source: togethercomputer/RedPajama-Data-V2. License: Apache 2.0 (data preparation code), Not specified (data) but see Common Crawl terms of use.

• Extracted from: Common Crawl.

• Description: "RedPajama-V2 is an open dataset for training large language models. The dataset includes over 100B text documents coming from 84 CommonCrawl snapshots and processed using the CCNet pipeline. Out of these, there are 30B documents in the corpus that additionally come with quality signals, and 20B documents that are deduplicated" (from GitHub). Most recent crawl for French data in the Lucie Training Dataset v1.1: 2023-14. (For more details on the time periods covered by crawls in this dataset see the composition details for French, German, Italian and Spanish.)

• Pre-processing and deduplication:

  • Url filtering:
    • Removing duplicate urls: urls were removed if their base domain overlapped with a dataset already in the Lucie Training Dataset (e.g., "theses.fr") in order to increase diversity of content (see code details).
    • Filtering certain toxic content: urls from a list of blacklisted content were removed (see code details).
    • Filtering by robots.txt files: we collect robots.txt and remove all documents for which CCBot is disallowed or for which we failed to collect information as of July 2024 in an effort to select data free from opt-out evidence according to the 4th article of the copyright European directive (2019).
  • Filtering: A series of filters were applied using quality signals already available in the dataset. This includes (see code details):
    • CCnet perplexity below 10 or above 1000
    • C4 filtering (including removal of documents that contain toxic words)
    • Gopher filtering and repetition removal
    • Redpajama document deduplication
  • Removal of personally identifying information (PII): email addresses and ip addresses were replaced with random addresses (see code details).
  • MinHash deduplication was performed on each snapshot and language independantly as proposed in FineWeb. For minhash configuration see code details.

  The Datatrove library was used to perform both filtering and deduplication stages.

• Citation: Together Computer (2023). "RedPajama-Data-v2: an Open Dataset with 30 Trillion Tokens for Training Large Language Models," GitHub.

STAC

• Source: STAC. License: CC BY-SA-NC 4.0.
• Description: A collection of multiparty chats from an online version of the game Settlers of Catan. The full STAC corpus contains annotations for discourse structure. We use only the text of the chats.
• Citation: Nicholas Asher, Julie Hunter, Mathieu Morey, Farah Benamara and Stergos Afantenos (2016). "Discourse structure and dialogue acts in multiparty dialogue: the STAC corpus," The Tenth International Conference on Language Resources and Evaluation (LREC 2016). European Language Resources Association, pp. 2721-2727.

TheStack (v1.2)

• Source: bigcode/the-stack-dedup. License: Other (mixture of copyleft licenses).
• Extracted from: GitHub via GHarchive. Mixed licenses for source.
• Description: "The Stack contains over 6TB of permissively-licensed source code files covering 358 programming languages. The dataset was created as part of the BigCode Project, an open scientific collaboration working on the responsible development of Large Language Models for Code (Code LLMs). The Stack serves as a pre-training dataset for Code LLMs, i.e., code-generating AI systems which enable the synthesis of programs from natural language descriptions as well as other from code snippets. This is the near-deduplicated version with 3TB data" (from the dataset card).
• Citation: Denis Kocetkov, Raymond Li, Loubna Ben Allal, Jia Li, Chenghao Mou, Carlos Muñoz Ferrandis, Yacine Jernite, Margaret Mitchell, Sean Hughes, Thomas Wolf, Dzmitry Bahdanau, Leandro von Werra and Harm de Vries (2022). "The Stack: 3 TB of permissively licensed source code," arxiv:2211.15533.

Theses

• Source: Corpus contributed by OpenLLM partners.
• Extracted from: theses.fr (License: Licence Ouverte / Open Licence version 2.0) and HAL (Open access).
• Description: A collection of doctoral theses published in France. Dataset containing text retrieved through OCR.
• Pre-processing:
  • Text cleaning:
    • Title pages about HAL, pages containing a significant fraction of control characters, and duplicate lines were removed (see code details).
    • Because the results of OCR on tables and graphics can give rise to garbage text, the text was cleaned by removing the most suspicious chunks. In particular, a chunk was removed if it was not detected as being written in French, English, Spanish, German or Italian, or if the perplexity of a CCNet Language Model on the chunk was higher than 2000 (see code details). The code to compute CCNET perplexity, parallelizing on parquet files, is available here.
  • Filtering: Texts with fewer than 1000 words or 10000 characters were removed (see code details).

Wikipedia, Wikisource, Wiktionary

• Source: Corpus contributed by LINAGORA Labs (OpenLLM-France). Also published here:
  • OpenLLM-France/wikipedia
  • OpenLLM-France/wikisource
  • OpenLLM-France/wiktionary
• Extracted from: Wikimedia dumps. License: GFDL/CC BY-SA.

YouTube

• Source: Corpus contributed by LINAGORA Labs and LeVoiceLab.
• Extracted from: YouTube.
• Description: French subtitles from videos published with permissive licenses on YouTube.
• Extraction pipeline description:
  • Searching for YouTube videos likely in French: Based on searches generated automatically from random sequences of words extracted from a corpus of French journalistic articles (initially obtained through a web-crawling tool applied to publicly accessible news and media sites such as Huffington Post, 20 Minutes, Le Parisien, Actu, Numerama, Slate, etc.).
    Selection of videos with subtitles labeled as "French," excluding those marked as "automatically generated."
    At this stage: 52,778 videos selected, corresponding to 10,654 hours of audio.
  • Selection of videos whose subtitle language classification confirms French with a certain confidence index:
    At this stage: 51,934 videos selected, corresponding to 10,425 hours of audio.
  • Selection of videos whose subtitles contain uppercase, lowercase, and punctuation marks:
    This step filters out automatically generated subtitles created with speech recognition tools.
    At this stage: 45,488 videos selected, corresponding to 8,904 hours of audio.
  • Extraction of audio tracks from the selected videos.
  • Automatic formatting of transcripts obtained from subtitles: Removal of emojis, sound event annotations in brackets (like "[Music]") and extra text such as "subtitled by XXX." (on last seconds of the video).
  • Selection of videos where an automatic speech recognition tool correctly transcribes the first 30 seconds with a minimum recall and precision rate:
    At this stage: 37,513 videos selected, corresponding to 7,541 hours of audio.
  • Realignment of the transcript: Ensuring accurate timestamps in the transcriptions based on the subtitles and excluding audios where alignment fails.
    At this stage: 36,618 videos selected, corresponding to 6,729 hours of audio.

Example use in Python

Load the dataset

Load and iterate over the full dataset using the datasets library:

from datasets import load_dataset

dataset = load_dataset("OpenLLM-France/Lucie-Training-Dataset", split="train", streaming=True)

for sample in dataset:
   
   text = sample["text"]

   # … do something with the text

Iterate over a subset

Several configurations are available to select a language, a source, or both, illustrated in the following examples.

The list of possible configurations can be obtained programmatically:

from datasets import load_dataset_builder

config_names = list(load_dataset_builder("OpenLLM-France/Lucie-Training-Dataset").builder_configs)

print(config_names)

['default', 'en', 'fr', 'de', 'es', 'it', 'de,fr', 'es,en', 'fr,en', 'it,en', 'natural', 'code', 'code-assembly', 'code-c', 'code-c#', 'code-c++', 'code-clojure', 'code-dart', 'code-elixir', 'code-erlang', 'code-fortran', 'code-go', 'code-haskell', 'code-java', 'code-javascript', 'code-julia', 'code-kotlin', 'code-lua', 'code-mathematica', 'code-matlab', 'code-ocaml', 'code-perl', 'code-php', 'code-python', 'code-r', 'code-racket', 'code-ruby', 'code-rust', 'code-scala', 'code-swift', 'code-tex', 'code-typescript', 'AmendementsParlement', 'AmericanStories', 'Claire', 'Claire-en', 'Claire-fr', 'CroissantAligned', 'DiscoursPublics', 'Europarl', 'Europarl-de', 'Europarl-en', 'Europarl-es', 'Europarl-fr', 'EuroparlAligned', 'EuroparlAligned-de,fr', 'EuroparlAligned-es,en', 'EuroparlAligned-fr,en', 'EuroparlAligned-it,en', 'Eurovoc', 'Eurovoc-de', 'Eurovoc-en', 'Eurovoc-es', 'Eurovoc-it', 'FineWebEdu', 'GallicaMonographies', 'GallicaPress', 'Gutenberg', 'Gutenberg-de', 'Gutenberg-en', 'Gutenberg-es', 'Gutenberg-fr', 'Gutenberg-it', 'HAL', 'InterventionsParlement', 'LEGI', 'MathPile', 'OpenData', 'OpenEdition', 'PeS2o', 'PeS2o-s2ag', 'PeS2o-s2orc', 'Pile', 'Pile-DM_Mathematics', 'Pile-FreeLaw', 'Pile-NIH_ExPorter', 'Pile-PhilPapers', 'Pile-StackExchange', 'Pile-USPTO_Backgrounds', 'Pile-Ubuntu_IRC', 'QuestionsEcritesParlement', 'RedPajama', 'RedPajama-de', 'RedPajama-es', 'RedPajama-fr', 'RedPajama-it', 'Stac', 'TheStack', 'Theses', 'Wikipedia', 'Wikipedia-de', 'Wikipedia-en', 'Wikipedia-es', 'Wikipedia-fr', 'Wikipedia-it', 'Wikisource', 'Wiktionary', 'YouTube']

Below are some examples of how to load data from different sources and in different languages.

Load data in French:

from datasets import load_dataset

kwargs = dict(split="train", streaming=True)

dataset = load_dataset("OpenLLM-France/Lucie-Training-Dataset", "fr", **kwargs)

Load data where French and English are aligned:

dataset = load_dataset("OpenLLM-France/Lucie-Training-Dataset", "fr,en", **kwargs)

Load data corresponding to files with programming languages:

dataset = load_dataset("OpenLLM-France/Lucie-Training-Dataset", "code", **kwargs)

Load data in Python:

dataset = load_dataset("OpenLLM-France/Lucie-Training-Dataset", "code-python", **kwargs)

Load data from Wikipedia (in all available languages):

dataset = load_dataset("OpenLLM-France/Lucie-Training-Dataset", "Wikipedia", **kwargs)

Load data from French pages of Wikipedia (wikipedia.fr):

dataset = load_dataset("OpenLLM-France/Lucie-Training-Dataset", "Wikipedia-fr", **kwargs)

Load the Pile dataset:

dataset = load_dataset("OpenLLM-France/Lucie-Training-Dataset", "Pile", **kwargs)

Load the subset "PhilPapers" from the Pile dataset:

dataset = load_dataset("OpenLLM-France/Lucie-Training-Dataset", "Pile-PhilPapers", **kwargs)

Load a specific version

You can load a specific version with the datasets Python package using the revision parameter of load_dataset(…):

from datasets import load_dataset

kwargs = dict(split="train", streaming=True)

name = None # or a configuration (e.g. "fr", "code-python", "Wikipedia-fr", "Pile-PhilPapers")

dataset = load_dataset("OpenLLM-France/Lucie-Training-Dataset", name, revision="v1.2", **kwargs)

Citation

When using the Lucie Training Dataset, please cite the following paper:

✍ Olivier Gouvert, Julie Hunter, Jérôme Louradour, Christophe Cérisara, Evan Dufraisse, Yaya Sy, Laura Rivière, Jean-Pierre Lorré (2025). The Lucie-7B LLM and the Lucie Training Dataset: Open resources for multilingual language generation. arxiv:2503.12294.

@misc{openllm2025lucie,
      title={The Lucie-7B LLM and the Lucie Training Dataset: Open resources for multilingual language generation}, 
      author={Olivier Gouvert and Julie Hunter and Jérôme Louradour and Christophe Cerisara and Evan Dufraisse and Yaya Sy and Laura Rivière and Jean-Pierre Lorré and OpenLLM-France community},
      year={2025},
      eprint={2503.12294},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2503.12294}, 
}

Acknowledgements

The Lucie Training Dataset was created by members of LINAGORA (Olivier Gouvert, Julie Hunter, Jérôme Louradour, Jean-Pierre Lorré) and the OpenLLM-France community.

We thank in particular Rachel Bawden (INRIA), Clément Bénesse (Opsci), Christophe Cérisara (LORIA), Evan Dufraisse (CEA List), Olivier Ferret (CEA List), Joöl Gombin (Opsci), Ismaïl Harrando (LINAGORA), Jordan Ricker (Opsci), Guokan Shang (MBZUAI), and Yaya Sy (LORIA) for their helpful input.

Data storage and significant parts of the data processing were made possible through the HPC resources from GENCI–IDRIS (Grant 2024-GC011015444).

Contact

contact@openllm-france.fr