Optimizing Large Language Model Performance
Content Analyst | E-Solutions | 10/2023 – 05/2024
Highlights -
Created clear rules and systems to analyze task features.
Utilized cognitive principles to teach new skills.
Conducted extensive secondary research, leveraging resources such as arXiv.
Designed content across various languages, including Python and JSON.
The Problem
The LLM lacked the ability to handle tasks that required complex instruction following, memory retention, and problem-solving. We used Reinforcement Learning from Human Feedback (RLHF) to guide the model’s learning, helping it align more closely with human expectations.
Goal
I worked with a team of content analysts, research scientists and linguists to fine-tune a GenAI LLM model to improve security and develop features.
Outcome
The model showed significant improvement in its ability to process complex instructions and follow multi-step tasks. By integrating RLHF during the solution phase, we ensured that the model could better align with human expectations. The adversarial prompts and cognitive insights we employed helped build a more robust system that was better equipped to handle tasks involving time and spatial awareness. Additionally, the annotation UI recommendations I made boosted annotator productivity by 25%, improving the training workflow.
