Toy demo: fine-tuning a 3.2M-param char transformer makes it forget Shakespeare (held-out numbers)

A 3.2M-parameter character-level transformer pre-trained on Tiny Shakespeare to a validation loss of 1.50 forgot its original training after 1,000 fine-tuning steps on Mrs Beeton's Book of Household Management (1861), with the prompt "ROMEO:" returning "377. INGREDIENTS.–4 oz. of salt" on text held out from both corpora. The author also froze the fine-tuned model, hashed its weights with SHA-256, prompted it two ways and re-hashed, finding different outputs but an identical hash. The experiment, presented as a toy demonstration of catastrophic forgetting, was used in a 5-minute explainer on training versus fine-tuning versus prompting.

I pre-trained a 3.2M-parameter character-level transformer on Tiny Shakespeare validation loss 1.50 , then fine-tuned it for 1,000 steps on Mrs Beeton’s Book of Household Management 1861 . Results on text held out from both: After fine-tuning, the prompt “ROMEO:” yields “377. INGREDIENTS.–4 oz. of salt”. A small, concrete example of forgetting. I also froze the fine-tuned model, hashed its weights SHA-256 , prompted it two ways and hashed again: different outputs, identical hash. I used both in a 5-minute explainer on training vs fine-tuning vs prompting for a general audience: https://www.youtube.com/watch?v=hwl4Ve0NBW8 https://www.youtube.com/watch?v=hwl4Ve0NBW8 It is a toy model and its output is nonsense with the right shape. Happy to answer questions about the setup. What would you use to demonstrate forgetting at this scale?