|
--- |
|
license: unknown |
|
language: |
|
- is |
|
pretty_name: IC3-v2 |
|
size_categories: |
|
- 1B<n<10B |
|
--- |
|
|
|
# Dataset Card for IC3-v2 |
|
|
|
<!-- Provide a quick summary of the dataset. --> |
|
The Icelandic Clean Crawled Corpus v2 (IC3-v2) is a collection of quality-filtered plaintext documents in Icelandic extracted from scraped websites with the `.is` top-level domain (TLD) in Common Crawl dumps between the years 2013 and 2023. |
|
The corpus contains about 1.3 billion words across almost 4 million documents. |
|
|
|
## Dataset Details |
|
We extract all WARC records matching the `.is` TLD from all available Common Crawl dumps as of the end of year 2023. Using a manually curated blacklist of domain names, we |
|
remove records from websites with gambling, pornography, and other illegal or harmful content. We extract the plaintext from the corresponding raw HTML |
|
code using `trafilatura` (with `jusText` as fallback). During the extraction process, `trafilatura` uses heuristics to provide fields such as `title`, `author` and `tags`, |
|
which we include as document-level metadata. We then apply various hand-crafted quality filters, similar to the Gopher rules and FineWeb filters, on the extracted |
|
output in an attempt at excluding low-quality documents like SEO product pages, documents with a high number of repeated sentences, etc. On the high-quality documents, |
|
we then run FAIR's `fasttext` language identification model and only keep documents with high proportion of Icelandic text. Finally, we run deduplication on the corpus. |
|
This we do by both running a sliding window of three sentences across each document and removing the span if it was previously seen in any document before, and also by |
|
performing document-wise exact string-matching deduplication across the corpus. |
|
|
|
Please note that even though the documents have been URL-filtered for harmful content on a best-effort basis, some may still remain. |
