---
language: pt
license: cc-by-4.0
multilinguality:
  - monolingual
task_categories:
  - text-generation
  - fill-mask
task_ids:
  - language-modeling
  - masked-language-modeling
size_categories:
  - 100M<n<1B
dataset_info:
  features:
  - name: url
    dtype: string
  - name: content_languages
    dtype: string
  - name: warc_filename
    dtype: string
  - name: warc_record_offset
    dtype: int64
  - name: warc_record_length
    dtype: int64
  - name: text
    dtype: string
  - name: crawl_timestamp
    dtype: string
  splits:
  - name: train
    num_bytes: 1087519823221
    num_examples: 342818651
  download_size: 1087713663056
  dataset_size: 1087519823221
---

# Canarim: A Large-Scale Dataset of Web Pages in the Portuguese Language

## Introduction

Canarim is a database encompassing over 342 million Portuguese language documents, sourced from multiple iterations of CommonCrawl. This nearly 1 terabyte database stands as one of the most extensive Portuguese language data collections available. It underwent initial deduplication using URLs, with plans for further text-based deduplication and filtering of potentially harmful content. The data, originally in HTML, has been converted to Markdown with the `Trafilatura` library to enhance readability and quality. Canarim is poised to be a crucial resource for NLP research, particularly in Portuguese language applications, filling the gap in large-scale, high-quality data for languages other than English.

## Dataset Structure

### Data Instances

An example looks as follows:

```json
{
  'url': '...',
  'content_languages': 'por',
  'warc_filename': 'crawl-data/CC-MAIN-2023-06/segments/1674764500041.18/warc/CC-MAIN-20230202200542-20230202230542-00352.warc.gz',
  'warc_record_offset': 971279893,
  'warc_record_length': 3873,
  'text': '...',
  'crawl_timestamp': '2023-02-02T20:28:21Z'
}
```

### Data Fields

- `url`: URL of the page
- `content_languages`: Language of the page
- `warc_filename`: Name of the WARC file
- `warc_record_offset`: Offset of the WARC record
- `warc_record_length`: Length of the WARC record
- `text`: Text of the page, in Markdown format
- `crawl_timestamp`: Timestamp of the crawl

WIP