About

Software engineer with a diverse background solving problems in multiple domains. Strong skills in machine learning and in AI Safety.

c1505 - Overview

Some of my recent public work

MMLU’s Moral Scenarios Benchmark Doesn’t Measure What You Think it Measures

• Report: MMLU’s Moral Scenarios Benchmark Doesn’t Measure What You Think it Measures — AI Alignment Forum

• About:

  • Results from the data from the leaderboard drove me to dive deeper into MMLU’s moral scenarios task. I found out that MMLU’s Moral Scenarios Benchmark isn't a useful measure of moral judgement.

  

Exploring the Characteristics of Large Language Models: An Interactive Portal for Analyzing 1000+ Open Source Models Across 57 Diverse Evaluation Tasks

• Deployed site: https://huggingface.co/spaces/CoreyMorris/MMLU-by-task-Leaderboard
• Code: https://huggingface.co/spaces/CoreyMorris/MMLU-by-task-Leaderboard/tree/main
• About:
  • Prior to creating this leaderboard, there was no publicly available leaderboard with a large amount of data from large language models broken down by the MMLU task
  • The evaluations were run by hugging face and the data comes from them. Full citations are on the site itself

• Medium Blog Post: Preliminary Analysis of MMLU-by-task: Insights from the Evaluation of Over 500 Open Source Models

Deep Reinforcement Learning Continuous control project from Udacity Deep Reinforcement Learning Nanodegree

• Report: https://github.com/c1505/deep-reinforcement-learning/blob/master/p2_continuous-control/Report.md
• Code: https://github.com/c1505/deep-reinforcement-learning/blob/master/p2_continuous-control/Continuous_Control.ipynb

First place on Hugging Face Leaderboard for the Deep-RL course on Pixelcopter-PLE-v0 environment (As of August 7, 2023. It may change)

• Model with video of it running: ****https://huggingface.co/CoreyMorris/ppo-rl-zoo-Pixelcopter-PLE-v0
• Link to leaderboard https://huggingface.co/spaces/huggingface-projects/Deep-Reinforcement-Learning-Leaderboard (You have to navigate to the Pixelcopter-PLE-v0 tab after clicking the link)