U-Tune: Fine-tuning
Gen AI Models for
Non-technical Users
IBM Research
UCSC Capstone Project
Team
Rocio Perez
Hebah Haque
Prajas Kadepurkar
Timeline
April 2024 - Dec 2024
Role
UX Researcher & Designer
Project Overview
The exponential development of generative artificial intelligence (gen AI) technology has created significant opportunities, yet these advancements remain largely inaccessible to non-technical users, exposing them to various risks.
​
To address this gap, we designed a tool to empower non-technical users to safely benefit from gen AI technologies, with the ability to understand, utilize, and fine-tune generative AI models.
Methods
Competitive Analysis
Examined 12 competitors through a SWOT analysis. Our focus was to identify gaps, strengths, weaknesses, and usability of these products. We identified a significant lack of resources for NTUs to effectively and safely fine-tune.
​
Literature Analysis
The literature focused on the safe use and transparency of AI (Trustworthy AI), and an emphasis on explainability.
​
Because of AI is rapidly evolving, we expanded our review to include informal literature, highlighted users’ pain points, such as high cognitive load, unintuitive user interfaces, and overwhelming parameters.
Expert Interviews
We conducted twelve semi-structured interviews with individuals of varying expertise in AI, gaining perspective on current fine-tuning
platforms, their shortcomings, and areas of opportunity from users with varying levels of fine-tuning knowledge.
Design Goals Informed by Desk Research
Analyzing the current commercial and academic landscapes helped us define our problem space: there are no adequate resources for NTUs to fine-tune. We translated the common themes from our interviews into design goals to address this problem:
​
-
Incorporate a trustworthy AI framework within our platform while simultaneously teaching users the basics of gen AI and fine-tuning
-
Lower the barrier to entry for fine-tuning by providing an intuitive, no-code interface that uses simple language
-
Encourage collaboration by creating spaces where users can interact and share ideas and experiences with one another.
The Solution: U-Tune
Key Features
Adjusting hyperparameters is often the most challenging part of fine-tuning. Most tools do not explain what each hyperparameter does, so users have to read through long, complicated documentation. Furthermore, most tools require users to change these settings using code, which can be difficult for NTUs.
U-Tune implements the following:​
-
Collapsable explanations with "Learn More" links
-
Short, simplified technical definitions
-
No-code adjustments using sliders
