Oskar Hallström ohallstrom

👋 Hello!

👨🏼‍💻 Currently aligning Large Language Models @ LightOn
- Conducted experiments with WSD-scheduler and novel positional weighting in loss on the Mamba Architecture
- Project lead in the creation of Alfred-40B-1023, using a custom extension method to obtain an 8K context length
- Responsible for reward modeling in the creation of Alfred-40B-0723 - a RLHF:ed version of Falcon-40B. Alfred was the second ever open source LLM aligned with PPO (Llama 2 being the first one).
🗒 MSc in Computer Science specialized in Machine Learning
- Master's Thesis in Synthetic Data Generation for Large Language Model Reward Modeling
- Included 65 credits at École Polytechnique Fédérale de Lausanne
👀 Interested in everything generative AI
📫 You can reach me on LinkedIn
🎸 Check out my Indie Band Billie Garlic