Success Story: ellamind summarizer
Our computationally intensive research project, “ellamind summarizer”, was completed in January. We would like to thank you very much for your support.
Here is a brief overview of the AI models we trained on the KISSKI cluster as part of the project:
sui-1: Grounded and Verifiable Long-Form Summarisation
sui-1 is a 24B LLM for long, verifiable summaries with inline citations, meaning that every statement can be traced back to a specific location in the source text. The model processes up to 100K tokens in a single pass and supports iterative processing for texts exceeding 2M tokens. A synthetic, multi-stage verified training pipeline (22k+ examples, 5 languages) achieves significantly better results than open baselines – even when compared to significantly larger models. The model was trained predominantly on German data and is already in use in the public sector.
propella-1: Multi-Property Document Annotation for LLM Data Curation at Scale
propella-1 is a family of small multilingual LLMs (0.6B/1.7B/4B) that annotate documents not with a single score, but along 18 properties across six categories as structured JSON. The models support 57 languages and, acting as an ‘annotator’, achieve high agreement with a commercial state-of-the-art LLM, with the 4B model outperforming larger general-purpose models. In collaboration with the OpenEuroLLM project, we are also publishing propella-annotations (3B+ annotations) as well as analyses that highlight differences in quality and content within common pre-training datasets that single-score approaches fail to capture. With propella-annotations, very large training datasets can be flexibly filtered, e.g. by topics such as medicine/healthcare or energy.
All models (sui-1 and propella-1) are released under the Apache 2.0 licence and are therefore freely available for commercial use.
If you have any questions or suggestions, please feel free to contact us at any time.