A 124M-parameter GPT trained from scratch on free Kaggle T4s across session resets.
Sources: Author blog post (5-6 Oct) · Note: This is a solo blog post and the results are self-reported. It is a pretraining demo, not a benchmark, and the author calls the result a repetitive base model. The reported held-out perplexity of 36.43 is on an in-domain split that is not fully decontaminated. CTN has not rerun it.