Datasets:

---

danish-foundation-models

/

swedish-dynaword

like 3

Follow

Danish Foundation Models 106

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

Some things to note:

• I used Liteparse (recommended by Yevhen) for pdf extraction. Extraction was conducted without OCR, but it seems that most documents don't need it. There are still some scanned documents that could benefit from it though.
• files longer than 100 pages or 100 mb were filtered out.
• I ran into out-of-memory errors for the native remove_duplicate_text function, so I modified it to use a hash instead of a set. I haven't included this in the PR, but it might be something to consider implementing in Dynaword.
• it might be worth using the same method to extract more data from Danish Cellar as well.

Thanks for the PR @gustavidun !

Some minor comments

1, I might be wrong on this, but can you check if LiteParse(..
max_pages=100, # to avoid memory crashes) actually filters the files longer then 100 pages, from my understanding it truncates them. Can you verify that

2. Perhaps add in limitations section that it may include some leakage of personal information

3. Can you please add line in the create.py how to run the script separately if others want to reproduce it (some thing like GIT_LFS_SKIP_SMUDGE=1 uv run data/cella/create.py

4. Maybe add link to the dataset into short description like Cellar

5. It is good to also add information about licence in the dataset card in the ## License Information section

1. You're right. I've changed the wording to reflect this. In the future, the max page limit can probably be removed, since the 100 mb check will catch the very heavy files anyway.
2. Wouldn't this data be GDPR compliant seeing as it is released and stored by the EU?
3. Added
4. Added
5. Added

In investigating 1., I discovered that there are quite a bit of long markdown tables in some of the heavier documents with mostly technical information (like chemical properties of meats). I've added a note about this in limitations, but can also filter out markdown tables if you prefer (would probably lose some natural language as well, though).

2. I don't think I would add it unless we know it is there (the metadata also partly covers this aspect)

I've added a note about this in limitations, but can also filter out markdown tables if you prefer (would probably lose some natural language as well, though).

I wouldn't. The annotation layer should handle this

I agree, probably they have covered it in the original release

I think it looks good, thanks @gustavidun for addressing the comments

I will merge it