Do you have any sources on this? I started looking around for pre-training, training and post-training impact of new input but didn’t find what I was looking for. In just my own experience with retraining (e.g. fine-tuning) pre-trained models, it seems to be pretty easy to add or remove data to get significantly different results than the original model.
LLMs acquire a wide range of abilities during pre-training, but aligning LLMs under Reinforcement Learning with Human Feedback (RLHF) can lead to forgetting pretrained abilities, which is also known as the alignment tax.
Do you have any sources on this? I started looking around for pre-training, training and post-training impact of new input but didn’t find what I was looking for. In just my own experience with retraining (e.g. fine-tuning) pre-trained models, it seems to be pretty easy to add or remove data to get significantly different results than the original model.
My go to source for the fact that LLM chatbots suck at writing reasoned replies is https://chatgpt.com/
It’s well-known folklore that reinforcement learning with human feedback (RLHF), the standard post-training paradigm, reduces “alignment,” the degree to which a pre-trained model has learned features of reality as it actually exists. Quoting from the abstract of the 2024 paper, Mitigating the Alignment Tax of RLHF (alternate link):