Joaoffg/ELM

Model Card for the Erasmian Language Model

ELM is a community driven large language model tailored to the research and education needs of Erasmus University (EUR, Netherlands) students and staff.

The model draws inspiration from ChatGPT and Llama in terms of architecture, but it aims to be privacy sensitive, environmentally conscious, and from and for the Erasmus community.

We hope that the ELM experience becomes a template for community driven, decentralized and purpuseful AI development and application.

November 2026 update: all personal data was removed from the training set and the model was retrained accordingly.

Model Details

Model Description

The underlying language model is trained and fine-tuned on academic outputs from Erasmus University, such as scientific papers or student theses; Training and fine-tuning the model is a joint effort of students and staff, transparent for all parties involved; The prompt-response examples used to fine tune the model come from students and staff, not crowdsourcing services; Defining what is the "better" model output also comes from the perspective of research and education. The true richness of ELM lies in the way its training data is generated. What is the "state-of-the-art" model may change quickly, but quality data will maintain its relevance and ensure that ELM and its future iterations serve the needs of the community that nurtured it.

• Developed by: João Gonçalves, Nick Jelicic
• Funded by [optional]: Convergence AI and Digitalization, Erasmus Trustfonds
• Model type: Llama-2 Instruct
• Language(s) (NLP): English, Dutch
• License: Llama2

Model Sources

• Repository: https://github.com/Joaoffg/ELM
• Paper: https://arxiv.org/abs/2408.06931
• Demo: https://huggingface.co/spaces/Joaoffg/Joaoffg-ELM

Copyright policy: ELM should be used exclusively for research purpuses and makes use of the Text Data Mining exception for scientific research. The size of the model should not allow it to reproduce the content that it was trained on, and this was verified in testing.

However, any authors of EUR Pure Research outputs or MA Theses in the EUR Thesis Repository who wish for their works to not be used in the model can contact ferreiragoncalves@eshcc.eur.nl to request that the model is retrained without their works.