Christian Boulangercmboulanger@sciences.social
Aug 24, 2026, 10:12 AM@Finto Looks really interesting! Can you say a bit more on what goes into the pipeline and what kind of metadata is generated? Thanks!
Replies
Finto-palveluFintoAug 24, 2026, 11:00 AM@cmboulanger , thank you! Currently only PDF inputs are supported, and the REST API output is a simple JSON object with the extracted metadata. However, in the future we would like to be able to process more types of input, such as EPUB files or images for example, and produce output according to user needs (MARC, BIBFRAME, Dublin Core...). We're exited to develop this further!
- Christian Boulangercmboulanger@sciences.socialAug 24, 2026, 11:11 AM
@Finto Would you have an example for the extracted metadata - it's probably in the docs but I didn't find it right away. Also, do you have a demo somewhere (e.g. HF), where one could upload PDFs and see the result? Thanks!
Aug 24, 2026, 1:31 PM@cmboulanger Here is the output of the GreyLitLM for a very recently published electronic dissertation from the University of Helsinki (source: https://helda.helsinki.fi/items/53bad33e-d00d-4aa6-b67b-8bcf50f4db73 )
- Christian Boulangercmboulanger@sciences.socialAug 24, 2026, 1:41 PM
- Christian Boulangercmboulanger@sciences.socialAug 24, 2026, 1:43 PM
@osma @Finto Maybe your platform can make use of the finetuned NuExtract3 Model for extracting Table of Content information that I am working on: https://sciences.social/@cmboulanger/117149370355856933
- Aug 24, 2026, 1:47 PM
@cmboulanger Indeed, it should be possible to add a table_of_contents field to the schema given to NuExtract3. Currently the schema/template is quite limited: https://github.com/NatLibFi/BIBRA/blob/2a14380b58678ab5c7a644b4673f8e97fc53d03e/bibra/backend/nuextract.py#L69-L83
We're not yet sure what would be the ideal metadata schema, as a lot of that is use case / context specific. But we will keep your request in mind!
- Aug 24, 2026, 11:21 AM
@cmboulanger @Finto I also wondered what kinds of documents it's designed to extract metadata from. I noticed greylit lms mentioned so I assume it's partly grey literature, but were other formats in scope? It looks very cool anyway!
- Aug 24, 2026, 1:25 PM
@mia We currently support two models - one is GreyLitLM trained on (Finnish, but multilingual) grey literature from institutional repositories. The other one is NuExtract by NuMind. In both cases the input is any kind of PDF file (we plan to support images in the future). The output metadata schema is quite simple, Dublin Core-ish, but we want to extend it so eventually it could produce MARC or BIBFRAME.
