Appendix 2: Public Sector AI Taxonomy and Map

• Pre-training generative AI systems is not the object of generative AI evaluation, which relies on expert and end-user acceptance in the specific context it is applied in, rather the general performance benchmarking undertaken in pre-training. Technically, pre-training (and prior fine-tuning) uses a combination of self-supervised, supervised and reinforcement learning. While fine-tuning may take place when the context has been established, this is only a means to improving the performance upon evaluation, not the end evaluation of the system per se.