Discussion #110 - Adding data from KB (historical letters and administrative publications)

#111
Danish Foundation Models org
edited Jan 26

Adding data from the Royal Danish library:
https://sprogteknologi.dk/organization/det-kongelige-bibliotek
Two separate datasets:

  • kb_historical_letters
  • kb_administrative_publication

Checklist

  • [X ] I have run the test suite using make test and all tests pass
  • [ X] I have added/changed a dataset:
    • [ X] I have updated descriptive statistics using make update-descriptive-statistics
    • [X ] I have bumped the version use make bump-version
  • [X ] If I have added a create.py script I have added the script dependencies required to run that script.
  • [X ] I have updated the CHANGELOG.md if appropriate

(The exact names are not set in stone in case someone has better suggestions)

KennethEnevoldsen changed pull request status to open
Danish Foundation Models org
edited Feb 2

Hi @balsab -great addition good to see the dataset get a big bump in terms of data!

structure of create.py

It seems like a lot of the functions that you implement are already in dynaword. I would just use those, e.g. see the example here:

https://huggingface.co/datasets/danish-foundation-models/danish-dynaword/blob/main/data/danske-taler/create.py
or
https://huggingface.co/datasets/danish-foundation-models/norwegian-dynaword/blob/main/data/wikipedia-nno/create.py

you also seems to have done a lot of undocumented work before the create.py. Download and everything should ideally be included in the create.py so that someone could rerun the whole thing from just the create.py.

documentation

Both datasheets are also missing some documentation:

  • e.g. link to the dataset sources (e.g. here)
  • I would also extend the description a bit more, currently it is very hard to know what these datasets include
  • there is for instance more information about the OCR in the original written text
  • the short description is a bit too long. These are used in the tables on the readme so please shorten them

Regarding filtering

  • what do we do non-dan languages? Do we remove those or only if they don't adhere to the stopword filter?
  • I would probably consider setting the boundary lower e.g. 0.1, but maybe you can extract 5 examples from the boundary region to see their quality?
Danish Foundation Models org

About filtering:

  • The texts which have both danish and non danish are kept in
0.19 examples ['https://hdl.handle.net/109.1.3/ac41dbe0-b003-4ac0-a891-3a2d6b3977ab', 'https://hdl.handle.net/109.1.3/edeb06cf-246d-4dac-b458-05528a23f176', 'https://hdl.handle.net/109.1.3/e1946aa3-bf9b-4fc4-ab4a-765b4a620e43', 'https://hdl.handle.net/109.1.3/7f957f6d-6b42-4f4c-aaf5-aaff1e0919b6', 'https://hdl.handle.net/109.1.3/7dd20ce9-129b-4576-8c7c-d1f833d24ed0', 'https://hdl.handle.net/109.1.3/3ec9cd99-f5be-4494-8c51-75a67f0df1bc', 'https://hdl.handle.net/109.1.3/c1faf125-be39-49aa-930e-70445ce31c8f', 'https://hdl.handle.net/109.1.3/75ba215e-bde7-4995-b9e0-e5117d7c985d', 'https://hdl.handle.net/109.1.3/5735ff89-637a-4f7b-9456-b0a3c65b0d94', 'https://hdl.handle.net/109.1.3/0a232ac4-401b-478a-ad8b-82cc4700c2f7', 'https://hdl.handle.net/109.1.3/da31f703-d775-4e92-a805-982e2ea934de', 'https://hdl.handle.net/109.3.1/70296086-ac47-4908-8b84-0e68170eac57', 'https://hdl.handle.net/109.3.1/9e1cd3dd-d9b0-4e90-9e37-45745bc3577f', 'https://hdl.handle.net/109.3.1/903ca6ec-4ded-4487-916c-739dc4ce1d86']
0.21 examples ['https://hdl.handle.net/109.1.3/5dc0d5be-9246-4411-aac5-8c6d09bb4fd8', 'https://hdl.handle.net/109.1.3/c277a140-14de-424e-b600-b0d18fe60f5f', 'https://hdl.handle.net/109.1.3/0e4ef64c-5a27-43cd-911a-4e85926e04f1', 'https://hdl.handle.net/109.1.3/0808ebf6-6882-4eff-b553-efb27ed56fa0', 'https://hdl.handle.net/109.1.3/196045d3-6ef4-47e5-8188-8f3a7a07d9a7', 'https://hdl.handle.net/109.1.3/792e90f1-09bd-43c3-8b6b-d5eca1667852', 'https://hdl.handle.net/109.1.3/4faeb8a8-71bd-4498-83aa-50173c57f2de', 'https://hdl.handle.net/109.1.3/cafc47d5-1517-4739-808f-bad9cf1a8c8d', 'https://hdl.handle.net/109.1.3/32dcf156-a1ac-4c1e-8b88-3d12a7295dc1', 'https://hdl.handle.net/109.1.3/87c7aba4-9178-4c55-86a0-6d5cbcf036dc', 'https://hdl.handle.net/109.1.3/ecf4f15b-019e-4446-824d-4950880252c3', 'https://hdl.handle.net/109.1.3/4e2918e4-98ce-4bd1-8492-109f35eef7fa', 'https://hdl.handle.net/109.1.3/1a4aa719-f040-4277-95ad-f82ca2f41d37', 'https://hdl.handle.net/109.1.3/98847674-7bb0-40ba-85bd-d98584f99dc7', 'https://hdl.handle.net/109.1.3/37e0b4ba-1e14-4a67-9eb4-a3aa7ce90dd7', 'https://hdl.handle.net/109.1.3/512e3dda-81a4-45fb-92f6-03d59fffe84e', 'https://hdl.handle.net/109.1.3/07f54945-d463-4b66-b811-3d2ff380854d', 'https://hdl.handle.net/109.1.3/1ed28550-2522-4f1f-b72d-9006c9e23088', 'https://hdl.handle.net/109.1.3/fbb38b2d-abe7-472e-af73-f3b173cdbf36', 'https://hdl.handle.net/109.1.3/4ceff8ce-9d86-4c66-89a6-513363351ff7', 'https://hdl.handle.net/109.1.3/519ed6d1-3541-4947-8acb-26d2bf579ebf', 'https://hdl.handle.net/109.1.3/98219a74-c981-4841-8181-a99adfef07e0', 'https://hdl.handle.net/109.3.1/08172cec-0cfd-4788-a8c9-94c5feb35002', 'https://hdl.handle.net/109.3.1/d66a842b-e12b-41ac-84dd-b0ec014827c0', 'https://hdl.handle.net/109.3.1/f1d17302-ddcf-4a65-b0ff-920f24411045', 'https://hdl.handle.net/109.3.1/7e53de31-56e9-42a2-8aa7-c4ccaac73be0', 'https://hdl.handle.net/109.3.1/f7b11d97-3d85-4ea9-b3e7-484051b7d257', 'https://hdl.handle.net/109.3.1/8ba49297-de12-4a4e-8c71-8019e8585749']

The 0.19 ones seem to be in danish, but mostly tables or lists of names. The 0.21 ones still seem to have a lot of numbers and lists but start to have more continous text chunks

Danish Foundation Models org

Ahh these are really good. I think we can keep 0.2, but add these to the docs

KennethEnevoldsen changed pull request status to merged

Sign up or log in to comment