---

📢 Sondage 2026 : Utilisation des datasets publiques de MediaTech

Vous utilisez ce dataset ou d’autres datasets de notre collection MediaTech ? Votre avis compte ! Aidez-nous à améliorer nos datasets publiques en répondant à ce sondage rapide (5 min) : 👉 https://grist.numerique.gouv.fr/o/albert/forms/gF4hLaq9VvUog6c5aVDuMw/11 Merci pour votre contribution ! 🙌

---

🇫🇷 French Consolidated Legislation Dataset (LEGI)

This dataset contains a semantic-ready and embedded version of the French full consolidated text of national legislation and regulations as published in the official LEGI database by Légifrance. The original data is downloaded from the dedicated DILA open data repository and is also published on data.gouv.fr.

The original and full LEGI dataset includes:

• All laws, codes, decrees, circulars, deliberations, decree-laws, ordinances etc... since 1945.
• A selection of consolidated ministerial orders (arrêtés).
• All official codes in force, and repealed codes.
• And more ...

In this version, all status are included :

• In force (VIGUEUR) : article that applies on the current date.
• Modified (MODIFIE) : content that has been modified, so not in force anymore
• Abrogated deferred (ABROGE_DIFF) : content still valid until the end date indicated.
• Abrogated (ABROGE) : article that is no longer in force as a result of explicit repeal by a text published in the Official Journal.
• Canceled (ANNULE) : article of the code cancelled by decision of the Council of State following an appeal.
• Disjointed (DISJOINT) : ‘separate’ article of the code; its provisions are no longer applied. This separation may not be permanent; its provisions may be reinstated by a new text. This legal status is specific to tax legislation.
• Modified stillborn (MODIFIE_MORT_NE) : article amended or repealed before the date set for its entry into force shall be deemed never to have had legal existence.
• Outdated (PERIME) : article subject to implicit repeal; this is the case, for example, with Article 39 octies of the General Tax Code.
• Transfered (TRANSFERE) : article whose provisions are reproduced under another article number. Example: Article L. 821-5-2 of the code de commerce has become Article L. 821-5-3 of the same code.
• In force deferred (VIGUEUR_DIFF) : article that comes into force at a later date. When this date is known, this deferred effective status is indicated.

To keep only certain statuses, you can check out the dedicated tutorial section below

Each article is chunked and vectorized using the BAAI/bge-m3 embedding model, enabling use in semantic search, retrieval-augmented generation (RAG), and legal research systems for example.

The dataset is splitted in subfolders by 'category' and 'code', so that the dataset is more usable for specifics use cases.

---

🗂️ Dataset Contents

The dataset is provided in Parquet format and includes the following columns:

Column Name	Type	Description
chunk_id	str	Unique chunk identifier.
doc_id	str	LEGI article source identifier.
chunk_index	int	Index of the chunk within its original document. Starting from 1.
chunk_xxh64	str	XXH64 hash of the chunk_text value.
nature	str	Nature of the text (e.g., Article).
category	str	Type of legislative document (e.g., LOI, DECRET, ARRETE, etc.).
ministry	str	Ministry responsible for the publication, if applicable.
status	str	Status of the article: VIGUEUR (in force), MODIFIE (modified) etc ...
title	str	Short title of the legislative document.
full_title	str	Full formal title of the text.
subtitles	str	Subtitle(s) indicating article grouping or section.
number	str	Article or section number.
start_date	str	Date the article entered into force (format: YYYY-MM-DD).
end_date	str	End of validity date (or 2999-01-01 if still in force).
nota	str	Additional notes, if present.
links	list[dict]	List of legal links associated with the text with metadata.
text	str	Textual content chunk of the article (or subsection).
chunk_text	str	Formatted full chunk including full_title, number, formatted subtitles and text.
embeddings_bge-m3	str	Embedding vector of chunk_text using BAAI/bge-m3, stored as JSON array string.

---

🛠️ Data Processing Methodology

📥 1. Field Extraction

The following fields were extracted and/or transformed from the original source:

• Basic fields:

  • doc_id, nature, category, ,ministry,status, title, full_title, subtitles, number, and start_date, end_date are extracted directly from the metadata of each article.

• Generated fields:

  • chunk_id: a generated unique identifier of the chunk, combining the doc_id and chunk_index.
  • chunk_index: is the index of the chunk of a same article. Each article has an unique doc_id.
  • chunk_xxh64: is the xxh64 hash of the chunk_text value. It is useful to determine if the chunk_text value has changed from a version to another.

• Textual fields:

  • nota: additional notes, also directly extracted from metadatas.
  • text: chunk of the main text content.
  • chunk_text: generated by concatenating full_title, article number, formatted subtitles and text.

• List-based fields:

  • links: list of legal relationships associated with the article, extracted from the LEGI metadata. Each element is a JSON object describing a link to another legal document (source or target), with fields such as:
    • title: official title of the related document.
    • doc_id / text_doc_id: LEGI or JORF identifier of the related document.
    • number / text_number: article or act number when applicable.
    • category: type of the related document (e.g. LOI, DECRET, etc.).
    • link_type: nature of the relationship (e.g. TXT_SOURCE, CITATION).
    • link_direction: direction of the relation, indicates whether the element carrying the link corresponds to the source or target of the link. (e.g. source when the current article cites another text, cible when it is cited by another text).
    • nor: NOR identifier of the related act, when available.
    • text_signature_date: signature date of the related text (YYYY-MM-DD), when available.

✂️ 2. Generation of 'chunk_text'

Each article is naturally isolated and has its own cid value. But in case the article is too long, we have to split it in several chunks. The Langchain's RecursiveCharacterTextSplitter function was used to make these chunks, which correspond to the text value. The parameters used are :

• chunk_size = 1024
• chunk_overlap = 0
• length_function = bge_m3_tokenizer

For each chunk (text), a chunk_text is constructed as follows:

"{full_title} - Article {number}
{formatted subtitles}
{text}"

🧠 3. Embeddings Generation

Each chunk_text was embedded using the BAAI/bge-m3 model. The resulting embedding vector is stored in the embeddings_bge-m3 column as a string, but can easily be parsed back into a list[float] or NumPy array.

🎓 Tutorials

🔄 1. The chunking doesn't fit your use case?

If you need to reconstitute the original, un-chunked dataset, you can follow this tutorial notebook available on our GitHub repository.

⚠️ The tutorial is only relevant for datasets that were chunked without overlap.

🤖 2. How to load MediaTech's datasets from Hugging Face and use them in a RAG pipeline ?

To learn how to load MediaTech's datasets from Hugging Face and integrate them into a Retrieval-Augmented Generation (RAG) pipeline, check out our step-by-step RAG tutorial available on our GitHub repository !

📌 3. Embedding Use Notice

⚠️ The embeddings_bge-m3 column is stored as a stringified list of floats (e.g., "[-0.03062629,-0.017049594,...]"). To use it as a vector, you need to parse it into a list of floats or NumPy array.

Using the datasets library:

import pandas as pd
import json
from datasets import load_dataset
# The Pyarrow library must be installed in your Python environment for this example. By doing => pip install pyarrow

dataset = load_dataset("AgentPublic/legi")
df = pd.DataFrame(dataset['train'])
df["embeddings_bge-m3"] = df["embeddings_bge-m3"].apply(json.loads)

Using downloaded local Parquet files:

import pandas as pd
import json
# The Pyarrow library must be installed in your Python environment for this example. By doing => pip install pyarrow

df = pd.read_parquet(path="legi-latest/") # Assuming that all parquet files are located into this folder
df["embeddings_bge-m3"] = df["embeddings_bge-m3"].apply(json.loads)

You can then use the dataframe as you wish, such as by inserting the data from the dataframe into the vector database of your choice.

🐱 GitHub repository :

The project MediaTech is open source ! You are free to contribute or see the complete code used to build the dataset by checking the GitHub repository

📚 Source & License

🔗 Source :

• DILA open data repository
• Data.gouv.fr : LEGI: Codes, lois et règlements consolidés

📄 Licence :

Open License (Etalab) — This dataset is publicly available and can be reused under the conditions of the Etalab open license.