Oskar Hallstrom

Oskar Hallstrom is a dedicated Machine Learning engineer with a deep focus on Generative AI, bringing innovative solutions to complex NLP challenges. Currently, he is serving as an ML/AI Research Intern at LightOn in Paris, where he plays a pivotal role in advancing large-scale experiments with Large Language Models (LLMs). In this position, Oskar is deeply involved in optimizing model adaptation techniques, exploring efficient fine-tuning methods, and pushing the boundaries of context length extension to enhance LLM performance across diverse applications.

One of Oskar’s key projects at LightOn involves leading the development of the OSS LLM Alfred-40B-1023, an instruction-tuned version of Falcon-40B. This project leverages custom adaptation methods to extend the model’s context window, enabling more coherent and context-aware interactions. His expertise in distributed systems and large-scale training infrastructure has been instrumental in managing the computational demands of these experiments, ensuring efficient resource utilization and scalability.

Additionally, Oskar has taken responsibility for Reward Modeling in the creation of OSS LLM Alfred-40B-0723, the first RLHF (Reinforcement Learning with Human Feedback) version of the model. This work involves integrating human-in-the-loop feedback to improve alignment and safety, showcasing his proficiency in reinforcement learning techniques and natural language processing. His skill set spans fine-tuning, generative adversarial networks, and advanced programming in Scala and SQL, positioning him as a versatile contributor to cutting-edge AI research. Through these projects, Oskar continues to deepen his expertise in generative models, distributed training, and NLP, making significant strides toward deploying more capable and aligned AI systems.