Missing articles in en dataset

#59
by davidmezzetti - opened

Hello,

I was excited to see an updated version of the Wikipedia dataset. Thank you for all the work in creating this dataset.

I was planning to use this as the new source for txtai-wikipedia. While I was able to rebuild with 20231101.en as the new input source, I noticed some inconsistencies when testing. It appears that some fairly popular articles are missing.

Here are a few examples that are missing.

It's possible this is a root issue with the upstream dataset. I used a modified version of the the olm/wikipedia project and those articles are in that dataset I manually built. I used the 20240101 dumps for this though, so again it's entirely possible it's an upstream issue.

I did notice one difference between the olm/wikipedia project and scripts found in the wikipedia project. olm/wikipedia removed this code.

-                    # Filter redirects.
-                    if raw_content is None or red_ is not None:
-                        beam.metrics.Metrics.counter(language, "filtered-redirects").inc()
-                        continue

Not sure if it's relevant but wanted to share.

Bumping this as i am encountering the same issue, here's a bunch more links that are not in the dataset that are in the lvl 5 vital articles: (there are 10k+ lvl 5 vital articles missing from the dump by my count)

'https://en.wikipedia.org/wiki/Rhizaria',
'https://en.wikipedia.org/wiki/Evert%20Taube',
'https://en.wikipedia.org/wiki/Pastel',
'https://en.wikipedia.org/wiki/Kurt%20G%C3%B6del',
'https://en.wikipedia.org/wiki/The%20Origins%20of%20Totalitarianism',
'https://en.wikipedia.org/wiki/Bed%20bug',
'https://en.wikipedia.org/wiki/History%20of%20Trinidad%20and%20Tobago',
'https://en.wikipedia.org/wiki/Billings,%20Montana',
'https://en.wikipedia.org/wiki/Tuba',
'https://en.wikipedia.org/wiki/Uncertainty%20principle',
'https://en.wikipedia.org/wiki/Birgit%20Nilsson',
'https://en.wikipedia.org/wiki/Gilbert%20N.%20Lewis',
'https://en.wikipedia.org/wiki/Human%20evolution',
'https://en.wikipedia.org/wiki/Sens%C5%8D-ji',
'https://en.wikipedia.org/wiki/Tasmania',
'https://en.wikipedia.org/wiki/Luc%20Montagnier',
'https://en.wikipedia.org/wiki/Indo-European%20languages',
'https://en.wikipedia.org/wiki/First%20Great%20Awakening',
'https://en.wikipedia.org/wiki/Poland',
'https://en.wikipedia.org/wiki/Split,%20Croatia',
'https://en.wikipedia.org/wiki/Primary%20school',
'https://en.wikipedia.org/wiki/Maimonides',
'https://en.wikipedia.org/wiki/Harry%20Cohn',
'https://en.wikipedia.org/wiki/Seychelles',
'https://en.wikipedia.org/wiki/New%20International%20Version',
'https://en.wikipedia.org/wiki/Islamic%20calendar',
'https://en.wikipedia.org/wiki/Surrealism'

can give the whole list for debugging if needed

More missing articles (from 20231101.en) that I encountered by chance: Galaxy, Sun, Moon, Earth, Food

I found a bunch of missing articles from 20231101.en , by trying to get Wikipedia articles from the Countries and Territories subsets of this dataset. By my count there are 150+ missing articles considering this dataset alone. Here's a list of69 page titles that I can't find in the dataset (I checked a few closely with the datasets' id and title fields, but they didn't exist):

{'Suriname', 'Russia', 'Equatorial Guinea',  'Maldives', 'Guinea',  'Pakistan', 'Ukraine', 'Lebanon', 'Republic of China', 'Somalia', 'Seychelles', 'Japan',  'Thailand', 'Portugal', 'Iran','Libya',  'United Kingdom', 'Malawi',  'Myanmar', 'Uganda',  'Uzbekistan', 'Marshall Islands', 'Mauritius',  'Serbia',  'Venezuela', 'Sudan', 'Tanzania', 'Gabon', 'Spain',  'Taiwan', 'Indonesia',  'Slovenia', 'Haiti', 'Democratic Republic of Congo', 'Mongolia', 'Philippines', 'Tonga', 'Tokelau', 'El Salvador', 'India', 'Kosovo', 'Vanuatu', 'Republic of Kosovo',  'Honduras', 'Iceland', 'Oman', 'Uruguay', 'Jamaica',  'Laos', 'Nigeria','Madagascar', 'Zambia', 'Zimbabwe', 'Tajikistan', "People's Republic of China", 'North Korea', 'Hong Kong', 'Singapore', 'Fiji', 'Italy', 'Greece', 'Eswatini', 'Montenegro', 'South Korea', 'Morocco', 'West Bank', 'Kazakhstan',  'Gaza Strip', 'Nicaragua'}

@albertvillanova we would really appreciate a response here, since missing such popular and critical articles is a major bug that negatively affects most research projects using this dataset. Thanks!

In the meantime, you can use one of the following methods that don't appear to have this issue:

In the meantime, you can use one of the following methods that don't appear to have this issue:

I tried the OLM dataset, but it's not a drop-in replacement because this Wikimedia one is just a download from HF servers, whereas the OLM one downloads the Dumps. The problem is that you get rate limited with the heavy parallelization of the script -- do you have a workaround for this?

(The rate limit is probably why this dataset is missing so many articles)

Sign up or log in to comment