WUR Library

News

ChatGPPeter: experimenting with Large Language Models

article_published_on_label
October 18, 2024

What started as a playful farewell gift turned into a fascinating and highly instructive exercise in exploring the potential of LLM technology for WUR Library. WUR Library's tech team developed ChatGPPeter, an internal, custom-built, language model based on the open-source Gemma 2 model.

ChatGPPeter

ChatGTPeter_screenshot.jfif

Generative Pre-trained Peter (ChatGPPeter) has been trained on the Library’s internal software documentation with guardrails to minimise hallucinations. During the development phase, the team opted to quickly test assumptions and make real-time adjustments to the model. We're still learning and focusing on the model’s multilingual capacity and its hallucinations. We've partly tackled the hallucinations by creating rules that prevent the model from answering if no context is available. We're also examining the requirements to minimise noise in the responses as well as its current limitations regarding answer structuring and accuracy. Finally, we're considering technical aspects, such as deployment, compute resource management, cost monitoring, and testing frameworks.

We named the tool ChatGPPeter in honour of our retiring colleague, Peter van Boheemen. He significantly contributed to developing Library products and services used not only in our library but also in other libraries across the Netherlands.

Learning and innovation

In experimenting with LLMs we want to comprehend both their potential and limitations. By exploring how tools like ChatGPPeter can be integrated into our library's production tech stack, we want develop solutions that streamline the workflow of WUR researchers and students. Ultimately, we aim to not only to incorporate these models into our library’s infrastructure but also to provide informed guidance on AI-related purchasing decisions for tools that support Library services.

11 questions on LLMs@WUR

Stay tuned as we continue to learn how LLMs can support your work. Please take a few minutes to answer these 11 questions on LLMs@WUR on this form.

Are you interested in ChatGPPeter? Let’s have a chat! Please contact mdt.library@wur.nl for more information.