karpathy/gpt2_1558M_final4_hf

This is a GPT-2 model trained in llm.c for 330K steps (of 1M batch size) on FineWeb-EDU.

A lot more detailed information is here: https://github.com/karpathy/llm.c/discussions/677 .

This model has a bit of a complicated history. I wanted to train it for 400K steps, i.e. (-x 400000), but it became unstable later in training and exploded around step 330K. Because I was losing my computing quota shortly, I decided to just rewind back to checkpoint 300K, and then instead of going all the way to 400K I started annealing linearly down to 330K. This went without incident and produced this model.

This is the longest I've trained a GPT-2 model for, and it reaches HellaSwag of 62.7 by the end.