UChicago's Ari Holtzman argues fine-tuning an LLM can degrade its performance on its target distribution due to pre-existing priors

He is seeking non-adversarial real-world examples of this degradation.