mrinaldi/UsenetArchiveIT

Usenet Archive IT Dataset 🇮🇹

Description

Dataset Content

This dataset contains Usenet posts from Italian language newsgroups belonging to the it and italia hierarchies. The data has been archived and converted to the Parquet format for easy processing. The only preprocessing conducted on the text was the removal of two conversations in which the VBS source code of the malicious script "ILOVEYOU" was present as it was shared by two users for didactical purposes on May, 5, 2000; these conversations were causing HF to flag the dataset as malicious.

This dataset contributes to the mii-community project, aimed at advancing the creation of Italian open-source Language Models (LLMs).🇮🇹 🤖

Descriptive Statistics

This dataset contains 85010057 posts from 11956999 threads in 539 newsgroups. Threads appear to have around 7 posts on average, with a median of 3 posts. The posts were created between 1995 and 2024. The text of all the posts together sum up to a total of 55885335313 characters, or approximately 10-20B tokens. The average length of the posts is 657 characters, and the median length is 380 characters.

Languages

The dataset should contain only Italian language posts, but it is possible that some posts are in other languages. The dataset has not been language filtered, as post were expected to be in Italian.

Dataset Structure

Features

Each record in the dataset has the following fields:

• title: The title of the post.
• author: The username of the author of the post.
• id: The unique identifier of the post.
• timestamp: The timestamp of the post.
• progressive_number: An integer identifying the thread number in the newsgroup.
• original_url: The URL of the original post on Google Groups.
• newsgroup: The name of the newsgroup the post belongs to.
• text: The text content of the post.

This repo contains the dataset in the Parquet format. The dataset is split into multiple Parquet files inside the parquet folder, each containing a portion of the records. The files are named usenet_converted_*.parquet, where * is a number indicating the order of the file. The original jsonl lines of the data are included as well as compressed bz2 files.

Example

Below is a complete JSON record for reference:

{"title": "Cerco Vespa", "author": "Castel", "id": 4423, "progressive_number": 1, "timestamp": "1998-06-26T09:00:00", "newsgroup": "it.annunci.usato", "original_url": "https://groups.google.com/g/it.annunci.usato/c/kKlb8hcSiX0/m/eXMsxa8SJ5gJ", "text": "Cerco vespa 125 o 50 in buono stato ma di vecchio modello, zone RE, PR, MO.\nGrazie in anticipo\nCastel\nlorenzo.c...@comune.re.it"}

Additional Information

Dataset Curators

This dataset was compiled and curated by Hugging Face users mrinaldi and ruggsea, as part of the mii-community dataset creation effort. The software used to download the post was written by mrinaldi and its source coude is reachable on github.

Dataset rationale

The dataset was created as part of a bigger effort to create various high-quality datasets of native Italian text, with the aim of aiding the development of Italian open-source LLMs.

The dataset is expected to be used for training and fine-tuning language models, as well as for other NLP tasks such as text classification, summarization, and translation. The column text contains the raw text of the posts, and the column newsgroup contains the name of the newsgroup the post belongs to, which can be used for classification tasks.

Usage

You can load the dataset directly from datasets using the load_dataset function. Here's an example:

from datasets import load_dataset

dataset = load_dataset("mrinaldi/UsenetArchiveIT")