README.md

---
annotations_creators:
- machine-generated
language:
- ca
language_creators:
- found
license:
- cc-by-4.0
multilinguality:
- monolingual
pretty_name: CATalog
size_categories:
- 10B<n<100B
source_datasets:
- extended|mc4
- extended|oscar
- extended|cawac
tags: []
task_categories:
- fill-mask
- text-generation
task_ids:
- masked-language-modeling
- slot-filling
- language-modeling
---

## Dataset Description

- **Homepage:** [Projecte AINA](https://projecteaina.cat/)
- **Paper:** ["A CURATEd CATalog: Rethinking the Extraction of Pretraining Corpora for Mid-Resourced Languages"]()
- **Point of Contact:** [Language Technologies Unit at Barcelona Supercomputing Center (BSC)](langtech@bsc.es)

### Dataset Summary

CATalog is a diverse, open-source Catalan corpus for language modelling. It consists of text documents from 26 different sources, including web crawling, news, forums, digital libraries and public institutions, totaling in 17.45 billion words.

### Supported Tasks and Leaderboards

- `Fill-Mask`
- `Text Generation`
- `other:Language-Modelling`: The dataset is suitable for training a model in Language Modelling, predicting the next word in a given context. Success is measured by achieving a low [Perplexity](https://huggingface.co/spaces/evaluate-metric/perplexity)score, indicating the model's proficiency in accurately predicting subsequent words.
- `other:Masked-Language-Modelling`: The dataset is designed for training models in Masked Language Modelling. This task involves predicting masked or hidden words within a sentence. Success is typically measured by achieving a high performance score, such as accuracy or [F1](https://huggingface.co/spaces/evaluate-metric/f1) score, on correctly predicting the masked tokens.

### Languages

This dataset is in Catalan (ca-ES). Coming from the web, some documents may contain other languages.

## Dataset Structure

### Data Instances

The dataset is provided in a CSV format, where each row corresponds to a single document and contains a document identifier, the text, a quality score, the strategy used to evaluate the document quality, languages, and a URL of the document, if available.

```
document    text    score   strategy    languages   url

```

### Data Fields

- `document`: text string containing the document identifier. Consists of the subdataset code, the part number and a document number.
- `text`: text string from the document, with paragraphs separated by two newlines escape sequences. It is meant to be used directly as input for language modelling.
- `score`: positive float number representing the document quality, ranging from 0, which represents the worst quality, to 1, the best quality.
- `strategy`: text string describing the type of evaluation applied to obtain the document score. generic_hard uses the heuristic evaluation from [CURATE](https://github.com/langtech-bsc/corpus-cleaner-v2) and perfect_score means that manual review was done and the highest score (1) is applied.
- `languages`: dictionary containing the document languages, with a percentage indicating the character ratio for each one.
- `url`: text string with the URL of the document, if available.

### Data Splits

We do not provide any canonical splits for CATalog.

## Dataset Creation

### Curation Rationale

CATalog is mainly built on filtered, non-overlapping versions of [CommonCrawl](https://commoncrawl.org/) snapshots and a smaller set of manually scored corpora from specific sources. We use the [CURATE](https://github.com/langtech-bsc/corpus-cleaner-v2) pipeline, which combines exact deduplication, language identification, and scoring heuristics.

In the design of CATalog, we adhere to the following values:

- (1) **Scale & Flexibility**. We intend to produce datasets that have a significant impact on the training of multilingual models in the range of 7B-180B parameters. Since Catalan is a medium-resource language and data acquisition is already a challenge, binary filtering will limit us in terms of the amount of data. By providing a score, we are able to easily filter the corpus according to our corpus according to our needs.
- (2) **Neutral scoring**. As opposed to ML-based filtering, we can use simple rules and heuristics to avoid introducing further bias into the model ([Dodge et al., 2021](https://arxiv.org/abs/2104.08758); [Welbl et al., 2021](https://arxiv.org/abs/2109.07445)). We only use [FastText](https://fasttext.cc/docs/en/language-identification.html) to reject documents in other languages.

During development, we performed comparative judgment experiments to evaluate the usefulness of the scoring from the [CURATE](https://github.com/langtech-bsc/corpus-cleaner-v2) pipeline, which appears in most documents in CATalog and is intended for further filtering and analysis. We found a moderate correlation between the score and the perceived quality of the text. Our main goal was to maximize the usability of the corpus without getting into a trade-off between quantity and quality.

### Source Data

#### Initial Data Collection and Normalization

We applied extensive data processing using our [CURATE](https://github.com/langtech-bsc/corpus-cleaner-v2) pipeline.

We first filter documents by their language content using [FastText](https://fasttext.cc/docs/en/language-identification.html). Only documents with at least 50% of characters in Catalan are kept. We then perform exact document deduplication. After this stage, we score each document with a tested set of 8 heuristic evaluators, inspired from other web filterings and from our own creation.

The following pre-existing datasets were used:
- [`OSCAR-2301`](https://huggingface.co/datasets/oscar-corpus/OSCAR-2301)
- [`OSCAR-2201`](https://huggingface.co/datasets/oscar-corpus/OSCAR-2201)
- [`CaText`](https://zenodo.org/records/5483031)
- [`MaCoCu-ca 1.0`](http://hdl.handle.net/11356/1837)
- [`caWaC`](https://huggingface.co/datasets/cawac)
- [`Colossal OSCAR 1.0`](https://huggingface.co/datasets/oscar-corpus/colossal-oscar-1.0)
- [`mC4`]({https://huggingface.co/datasets/mc4)

#### Who are the source language producers?

Apart from the pre-existing datasets, all of them coming from [CommonCrawl](https://commoncrawl.org/) dumps, the following
sources provided their data on Open Data Agreements:
- ## Media Groups
- [`IB3`](https://ib3.org/)
- [`Grup El Món`](https://grupmon.cat/)
- [`Vilaweb`](https://www.vilaweb.cat/)
- [`Nació Digita`](https://www.naciodigital.cat/)
- [`ACN`](https://www.acn.cat/)
- [`Racó Català`](https://www.racocatala.cat/)
- [`Aquí Berguedà`](https://www.aquibergueda.cat/)
- ## Academic & Book Repositories
- [`Tesis Doctorals en Xarxa`](https://www.tesisenred.net/)
- [`Wikipedia`](https://ca.wikipedia.org/)
- [`Project Gutenberg`](https://www.gutenberg.org/)
- ## Government Institutions
- [`Valencian Parliament`](https://www.cortsvalencianes.es/)
- [`Diari Oficial de la Generalitat Valenciana`](https://dogv.gva.es/)
- [`Butlletí Oficial de la Universitat d'Alacant`](https://www.boua.ua.es/)

### Annotations

[N/A]

#### Annotation process

[N/A]

#### Who are the annotators?

[N/A]

### Personal and Sensitive Information

Being mostly constructed from Common Crawl, personal and sensitive information might be present. 
This must be considered before training deep learning models with CATalog, specially in the case of text-generation models.

## Considerations for Using the Data

### Social Impact of Dataset

CATalog promotes the Catalan language in the NLP field, enabling development of advanced applications and chatbots tailored to Catalan speakers, while improving access to information for better community understanding. However, most of the sources in the dataset are web-scraped, which may bring in biases and privacy issues, risking biased outcomes and potential misuse. Additionally, it might overlook the voices of low-resource communities, amplifying existing disparities in representation.

Given that Catalan is a mid-resourced language with representation in digital sources, this dataset becomes crucial for building inclusive NLP applications. It addresses the language's underrepresentation, empowering communities with improved access to information in their native language. However, careful consideration of potential biases and privacy issues is essential to ensure responsible and equitable technology use.

### Discussion of Biases

Web-crawled content is over-represented with standard language varieties, impacting language model performance for minority languages. Language diversity in data is crucial to avoid bias, especially in encoding non-standard dialects, preventing the exclusion of demographic groups. Our corpus primarily focuses on Central Catalan, but we actively include Valencian and Balearic Catalan, along with diverse sociolects from platforms like Racó Català Fòrums, aiming for a more representative dataset. Despite legal uncertainties in web-scraped data, we prioritize permissive licenses and privacy protection measures, acknowledging the challenges posed by personally identifiable information (PII) within large-scale datasets. Our ongoing efforts aim to address privacy concerns and contribute to a more inclusive linguistic dataset.

### Other Known Limitations

[N/A]

## Additional Information

### Dataset Curators

[Language Technologies Unit](langtech@bsc.es) at the Barcelona Supercomputing Center (BSC).

This work was funded by the [Departament de la Vicepresidència i de Polítiques Digitals i Territori de la Generalitat de Catalunya](https://politiquesdigitals.gencat.cat/ca/inici/) within the framework of [Projecte AINA](https://politiquesdigitals.gencat.cat/ca/economia/catalonia-ai/aina).

### Licensing Information

[Creative Commons Attribution 4.0 International](https://creativecommons.org/licenses/by/4.0/)

### Citation Information

[N/A]

### Contributions

[N/A]