Erasmus University Rotterdam and open-access database CORE (COnnecting REpositories) are jointly launching SHARE, a new series of AI models specially designed for the social sciences and humanities (SSH). These models are intended to help researchers, lecturers and students explore and develop new perspectives in research, teaching materials and discussions. The SHARE project team consists of João Gonçalves (ESHCC), Sonia de Jager (ESPhil), and Nick Jelicic (Future Library Lab).
SHARE (Social-Humanities AI for Research and Education) addresses a critical gap in the LLM landscape. LLM’s are Large Language Models, where AI chatbots like ChatGPT are built on. Most language models are trained on general web data or technical sources dominated by the STEM disciplines (Science, Technology, Engineering and Mathematics). As a result, they are often unable to understand or represent the nuances, context and complex theories in disciplines such as sociology, history, philosophy and economics.
Interdisciplinarity and the ethics of collecting and sharing data
The team decided to explicitly approach open-source libraries and enter conversations with data partners, rather than simply gather available data. This resulted in the collaboration with CORE, as well as with Open Humanities Press and the University of California Library. In addition to open-access research papers from these databases, the project team uses specialised sources such as Wikipedia, Project Gutenberg and the PeS2o dataset to train their models.
SHARE is being trained to move beyond generating superficial texts. The idea is for it to become a model with specialized linguistic and conceptual skills, capable of contextualizing scientific argumentation, interpreting complex theories and supporting researchers in new ways.
João Gonçalves, associate professor of AI and Digitalization at Erasmus University Rotterdam and principal investigator of the project, is enthusiastic: ‘With SHARE, we are creating a research tool for the social sciences and humanities. We are demonstrating what is possible in an area that has traditionally been neglected by academic AI models.’
Research assistant from lab to lecture hall
Gonçalves emphasises that SHARE is designed as an extended-cognition partner to encourage deeper thinking, with practical applications for both staff and students.
‘A researcher can use the model to brainstorm new research ideas or to get feedback on ongoing research,’ explains Gonçalves. The team also sees an important role for the tool in education. ‘A lecturer can have students prepare arguments using the model for a classroom discussion. This enriches the debate with perspectives that might not otherwise have been considered.’
'Our goal is to move from simple information gathering to a tool that can assist in conceptual analysis and stimulate new interdisciplinary discoveries.'
The project is developing two versions of the AI model. The larger version, with 14 billion parameters (a measure of the model's complexity and ‘thinking power’), is currently in the training phase in the Dutch supercomputer Snellius, and a smaller version with 4 billion parameters is already being tested.
‘Our goal is to move from simple information gathering to a tool that can assist in conceptual analysis and stimulate new interdisciplinary discoveries. This shows what can be achieved when we combine high-quality, open scientific data with advanced AI techniques, and the initial results of the 4B parameter model look promising,’ says Gonçalves. Additionally, as part of Sonia de Jager’s research on the topic, the team is also working on conceptualizing and eventually prototyping novel user interface possibilities that go beyond the linear chatbot design, hoping to offer novel ways to interact with LLMs than the standard two-way conversation model.
Support from international and local grants
Building a digital tool of this scale requires enormous computing power. The technical backbone of the project was made possible by a prestigious NVIDIA Academic Grant, which provided the team with 20,000 hours of GPU computing power. SHARE's technology is based on the Phi model architecture (developed by Microsoft for small language models) and uses advanced training methods to ensure efficiency and scalability.
The training process for the 14B model uses techniques such as Fully Sharded Data Parallelism (FSDP), the Triton Kernel and FlashAttention, as described in the project's open-source training script. In addition to NVIDIA, the training of the 14B model is being supported by SURF and NWO through compute grants for the Snellius supercomputer.
Open science
True to the university's commitment to societal impact, this technology will not remain behind closed doors. The SHARE models, the code used to build them, and a detailed research paper will be made public on the Hugging Face platform as soon as the current testing for usability, safety, privacy and bias is concluded.
The collaboration between Erasmus University Rotterdam and partners demonstrates how sustainable, open scientific infrastructures such as CORE are driving the next generation of domain-specific AI tools. By building a tool tailored to the unique needs of academics in the social sciences and humanities, the SHARE project promises to accelerate research, improve education and open up new avenues for understanding the human condition.
- Researcher
- Researcher
- Researcher
- More information
- Related content
