Datasets:

wikimedia

/

wikisource

like 69

It appears that there are (yet unquantified, but noticeable) records that simply provide a list of links to the source material instead of the text of the source itself. For example:

import datasets
ds = datasets.load_dataset("wikimedia/wikisource", "20231201.en")
for sample in ds["train"]:
   if sample['id'] == "1":  # Gettysburg Address
      text = sample["text"]
      break

print(text)

Prints the list of links to various versions of the Gettysburg Address:

Gettysburg Address (1863), Nicolay draft
 Gettysburg Address (1863), Hay draft
 Gettysburg Address (1864), Everett copy
 Gettysburg Address (1864), Bancroft copy
 Gettysburg Address (1864), Bliss copy
 Gettysburg Address (1922), inscribed on the southern interior wall at the Lincoln Memorial

Works about the Gettysburg Address
 Inauguration of the Gettysburg Cemetery (1863), report in the Daily Illinois State Journal

American Civil War
American speeches

This is in line with the wikisource article of the Gettysburg Address but is unhelpful if the intended use case is providing natural language text for language modeling.
I would prefer that these links be followed such that the plaintext could be provided in the wikisource dump; alternatively, it may be worth adding some feature to identify these "listicles" such that they could be easily identified/excluded?

This may be an offshoot of the discussion on cleaning here, or the referenced discussion from that issue (wikimedia/wikipedia #51)

Thank you!