A couple of weeks ago, a group of talented students from the University of Limerick’s Immersive Software Engineering program showcased their skills at the Zerve Hackathon. The course was divided into teams, each demonstrating creativity and technical expertise, with one team ultimately taking home the top prize.

The Challenge

The hackathon, hosted by Zerve, asked participants to apply a Large Language Model (LLM) of their choice to a real or hypothetical use case. Points were awarded for enhancing the performance of the model using Retrieval-Augmented Generation (RAG) or fine-tuning with new data. Teams had to demonstrate their solution with a working prototype or script.

The Winning Solution: Dionysus - The God of Entertainment

The winning team impressed judges with Dionysus, a project designed to extract features from media and identify the types of audiences most likely to resonate with that content.

1. Initial Approach with Claude3

• Started with Claude3 for recommendations based on user input, e.g., fans of “Dune.”
• Results were repetitive and uninspired, highlighting the need for refinement.

2. Implementing RAG

• Integrated APIs for trending movies, TV shows, books, and podcasts.
• Used this external data to enrich prompts for more relevant recommendations.

3. Generating Enhanced Recommendations

• Created prompts with genres and features to get more diverse outputs.
• Delivered improved and engaging media recommendations related to “Dune.”

4. Providing Explanations

• Fine-tuned a Microsoft model with GPUs from Zerve.
• Generated clear explanations for recommendations, enhancing trust and usability.

5. Ensuring Data Security

• Hosted the fine-tuned model directly within Zerve.
• Kept user prompts private and protected from third-party access.

6. Visual Enhancements

• Used Claude 3 with an image generation model to create visuals for recommendations.
• Added a creative and interactive element to the project.

Enhancing LLM Performance

The Dionysus project addressed core LLM limitations:

• Outdated data: Used APIs for current, trending content.
• Narrow focus: Generated results based on features, not just titles.
• Hallucinations: Grounded outputs in accurate external data.
• Inconsistent formatting: Standardized and reliable outputs.
• Single-task limitation: Ran parallel processes for text and images.

Conclusion

We were impressed by the innovation across all teams at the Zerve Hackathon. Dionysus stood out for combining technical skill with creativity, enhancing LLMs with real-time data, RAG, and visual outputs. See their winning submission here.

The students’ performance highlights the next generation of innovators in AI, and we’re excited to watch their contributions unfold.

FAQs

What is the Zerve Hackathon?

The Zerve Hackathon is a competition where participants use Zerve to build projects with Large Language Models (LLMs), focusing on enhancing model performance with techniques like Retrieval-Augmented Generation (RAG) or fine-tuning.

Who won the Zerve Hackathon?

A team from the University of Limerick’s Immersive Software Engineering program won with their project Dionysus, which created audience-based media recommendations using RAG and fine-tuning.

How did the winning team improve LLM recommendations?

They combined Claude3 with real-time data from APIs, applied RAG to enrich prompts, and fine-tuned a model to generate clear explanations, producing accurate and engaging recommendations.

Why was Zerve important for the hackathon projects?

Zerve allowed teams to fine-tune and run models securely within its platform, ensuring data privacy while also providing access to GPUs for scalable workloads.

What skills did students gain from the Zerve Hackathon?

Students gained hands-on experience with LLMs, RAG, data integration from APIs, fine-tuning models, ensuring data security, and presenting AI solutions with both technical and visual enhancements.