Oskar Hallström ohallstrom

👋 Hello!

👨🏼‍💻 Currently aligning Large Language Models @ LightOn
- Debugged very odd bugs, such as when I found torch.randperm to be not actually random
- Contributed to the model arch and training of the SOTA encoder ModernBERT
- Did a small side mission experimenting with a novel positional weighting in pretraining
- Project lead in the creation of Alfred-40B-1023, using a custom extension method to obtain an extended context length
- Responsible for reward modeling in the creation of Alfred-40B-0723 - a RLHF:ed version of Falcon-40B. Alfred was one of the first ever open LLMs aligned with PPO.
🗒 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 or X/Twitter
🎸 Check out my Indie Band Billie Garlic