Faculty Mentor
Joshua Farara
Location
Russell Union Ballroom
Type of Research
Proposed
Session Format
Poster Presentation
College
Allen E. Paulson College of Engineering & Computing
Department
Department of Computer Science
Abstract
Numerous artificial intelligence GPT models are trained on various language dataset types, including ethical and mental health dialogue, to produce ethically coherent responses, and these datasets are opened and merged together via the efficient processes of data mining and data merging. To explore how companies such as OpenAI and Microsoft achieve these goals, we implemented our own custom-built GPT, named AkiraGPT, which has recently and historically had issues generating ethical responses, simulating self awareness, and has sometimes generated incomprehensible, contextually irrelevant responses. In addition to that, AkiraGPT would sometimes have it's responses be "cut off."
Our goal with our research is to deploy the mental health dataset we have merged from other datasets, which we have mined, into a fresh experimental version of AkiraGPT to retrain the new model path with appropriately adjusted back propagation handles such as weights, so that the model could hopefully tokenize inputs and generate ethically coherent, contextually relevant responses via attention mask layering.
Program Description
.
DOI
10.20429/GS4.2026.005
Start Date
4-23-2026 10:00 AM
End Date
4-23-2026 12:00 PM
Recommended Citation
Moore, Rocco and Farara, Joshua, "Project Akira: Data Mining and Merging, and training a Custom-Built GPT to produce Therapist-Like, Ethically Coherent Responses" (2026). GS4 Student Scholars Symposium. 67.
https://digitalcommons.georgiasouthern.edu/research_symposium/2026/2026/67
Project Akira: Data Mining and Merging, and training a Custom-Built GPT to produce Therapist-Like, Ethically Coherent Responses
Russell Union Ballroom
Numerous artificial intelligence GPT models are trained on various language dataset types, including ethical and mental health dialogue, to produce ethically coherent responses, and these datasets are opened and merged together via the efficient processes of data mining and data merging. To explore how companies such as OpenAI and Microsoft achieve these goals, we implemented our own custom-built GPT, named AkiraGPT, which has recently and historically had issues generating ethical responses, simulating self awareness, and has sometimes generated incomprehensible, contextually irrelevant responses. In addition to that, AkiraGPT would sometimes have it's responses be "cut off."
Our goal with our research is to deploy the mental health dataset we have merged from other datasets, which we have mined, into a fresh experimental version of AkiraGPT to retrain the new model path with appropriately adjusted back propagation handles such as weights, so that the model could hopefully tokenize inputs and generate ethically coherent, contextually relevant responses via attention mask layering.
