malteos

Contributions are always welcome! I'll try to update the API accordingly within the next week.

The decision content which is currently available via the API does not contain any annotations. Thus, it should not be affected by the reference extraction bug. The API serializer returns...

Hi @fchrubasik & @dennlinger thanks again for your contribution! The last months have been really busy over here so I only today managed to finally deploy your changes to production....

The case mentioned in the issue seems to have all reference correct ( https://de.openlegaldata.io/case/bag-2019-07-11-6-azr-4017 ). Do you have an example for still misaligend references?

OK. Then let's reopen this one.

There is already a PyTorch implementation from @lucidrains (no trained weights though): https://github.com/lucidrains/RETRO-pytorch

KNN is not performed on the full S2ORC but only on the pool of candidate papers (~680k papers). The candidate papers come from SPECTER's training corpus (see Appendix C). >...

You could try something like this: ```python from transformers import AutoModelForSequenceClassification model = AutoModelForSequenceClassification.from_pretrained("malteos/scincl") ``` However, SciNCL is trained to produce embeddings and that might make it suboptimal compared to...

Is somebody already actively working on this?

My recommendation would be to take a math-optimized LLM, like [Llemma](https://arxiv.org/abs/2310.10631), feed the formulas through the model, take the embeddings and do a k-nearest neighbor search in the embedding space....