Our Open Source Contributions
Over the years, we have published parts of our code and our AI models. We have summarised these contributions on this page.
Multilingual Passage Reranking Model
- This model is a reranking model trained by us, which can evaluate content according to various factors on an LLM basis.
- It has been rated as one of the most efficient of its kind in various scientific papers:
- On cross-lingual retrieval with multilingual text encoders
- Improving NLP Model Performance on Small Educational Data Sets Using Self-Augmentation
- Query-Based Retrieval of German Regulatory Documents for Internal Auditing Purposes
- Better Retrieval May Not Lead to Better Question Answering
- Answer Similarity Grouping and Diversification in Question Answering Systems
- On cross-lingual retrieval with multilingual text encoders
- It now has a high six-digit number of downloads
High quality, German training data set
- High-quality data sets are required to train AI models and avoid certain biases.
- Together with Philipp May, T-Systems, we have defined this training dataset to train our LLMs.
- When we train AI models, we do not use customer data, but general, public data – such as from this training dataset.
German Electra Model
- Together with Philip May, T-Systems, we have published an improved Electra model, which was trained with 1.5 million steps.
- At the time of publication in August 2020, it was briefly the most efficient German-language LLM.
- Of course, it is now far from state of the art. But it shows how early we started working with large language models and that we have own expertise in this technology.