Try Llama-3 in a Colab Notebook

Made a Colab notebook to finetune Llama-3 2x faster and use 70% less VRAM on free T4 GPUs, but you can also directly inference with it etc!

15 trillion tokens is yikes!