Uploaded March 2024 | Updated September 2026, 2 weeks ago
The recent and sudden rise of Large Language Models (LLMs) offers great opportunities for accelerated research and development. Researchers have identified many scientific use-cases where LLMs can speed up the analysis, knowledge retrieval, and ultimately the execution of their research. Yet, despite the promise and excitement, there lies a fundamental issue with its use: the reluctance (and indeed inability) for organisations to use such technologies due to ethical, legal and privacy concerns.
Simply put, organisations have discouraged (and sometimes even barred) their staff from using cloud-based tools such as OpenAI’s ChatGPT. This is due to concerns of data leakage, where employees may inadvertently share sensitive data which is then stored with unauthorised third parties and subsequently ingested permanently into the models. There are also ethics and privacy concerns relating to how these LLMs were trained, including data that has no legal provenance (non-consent) and personally identifiable information.
Potential solutions include so called Private/Local LLMs, which seek to train and utilise such models in local environments, such as an organisation’s on-premises cloud or even on personal laptops. There exists a wide range of open-source projects dedicated to such efforts, with the ultimate aim being to train a usable LLM but with datasets that has been vetted and approved.
The benefit of being to efficiently train and utilize a controlled model is two-fold. Firstly, we can repurpose their generative/transformative aspects into information retrieval systems instead of simple (albeit impressive) chat apps. Secondly, we can exercise control over the input materials, which bestows us ultimate agency over how the model is trained. This is an important step in assuaging the aforementioned legal and ethical concerns.
Consequently, this allows us to train secure local models that have specific express purpose or functionality. LLMs of this nature are trained only to a specific semantic sufficiency (and no more) to assist researchers within their domains. This is done by applying parameter fine-tuning procedures, and limiting the training data to a specific domains and subsets (e.g., only text from peer-reviewed climate change publications).
CSIRO Energy, in collaboration with CSIRO Environment, has identified one use-case where such local LLMs can be trained on a large corpus of commercial building datasets and reference manuals. This allows a researcher to extract pertinent information about the equipment in the building and generate accurate semantic models. Thus, we can shorten the process to onboard a building from months to days.
Thus, this project aims to investigate the myriad of open-source models available and deploy a selected few within Pawsey’s protective and secure infrastructure. Subsequently, we will attempt to train, and apply inference time techniques to, such models with selected domain specific datasets, evaluate their performance as expert knowledge systems, and measure their utility for 1-2 selected use-cases.
The recent and sudden rise of Large Language Models (LLMs) offers great opportunities for accelerated research and development. Researchers have identified many scientific use-cases where LLMs can speed up the analysis, knowledge retrieval, and ultimately the execution of their research. Yet, despite the promise and excitement, there lies a fundamental issue with its use: the reluctance (and indeed inability) for organisations to use such technologies due to ethical, legal and privacy concerns.
Simply put, organisations have discouraged (and sometimes even barred) their staff from using cloud-based tools such as OpenAI’s ChatGPT. This is due to concerns of data leakage, where employees may inadvertently share sensitive data which is then stored with unauthorised third parties and subsequently ingested permanently into the models. There are also ethics and privacy concerns relating to how these LLMs were trained, including data that has no legal provenance (non-consent) and personally identifiable information.
Potential solutions include so called Private/Local LLMs, which seek to train and utilise such models in local environments, such as an organisation’s on-premises cloud or even on personal laptops. There exists a wide range of open-source projects dedicated to such efforts, with the ultimate aim being to train a usable LLM but with datasets that has been vetted and approved.
The benefit of being to efficiently train and utilize a controlled model is two-fold. Firstly, we can repurpose their generative/transformative aspects into information retrieval systems instead of simple (albeit impressive) chat apps. Secondly, we can exercise control over the input materials, which bestows us ultimate agency over how the model is trained. This is an important step in assuaging the aforementioned legal and ethical concerns.
Consequently, this allows us to train secure local models that have specific express purpose or functionality. LLMs of this nature are trained only to a specific semantic sufficiency (and no more) to assist researchers within their domains. This is done by applying parameter fine-tuning procedures, and limiting the training data to a specific domains and subsets (e.g., only text from peer-reviewed climate change publications).
CSIRO Energy, in collaboration with CSIRO Environment, has identified one use-case where such local LLMs can be trained on a large corpus of commercial building datasets and reference manuals. This allows a researcher to extract pertinent information about the equipment in the building and generate accurate semantic models. Thus, we can shorten the process to onboard a building from months to days.
Thus, this project aims to investigate the myriad of open-source models available and deploy a selected few within Pawsey’s protective and secure infrastructure. Subsequently, we will attempt to train, and apply inference time techniques to, such models with selected domain specific datasets, evaluate their performance as expert knowledge systems, and measure their utility for 1-2 selected use-cases.
