A Reproduction of
OpenLLaMA using 128 H100 GPUs in Bfloat16.
The pretrain data consists of Falcon, Starcoder, and the wikipedia, arxiv, books, stackexchange from RedPajama. In total, this encompassed nearly 1 trillion tokens.
The model was trained over a single epoch, incorporating 2000 warm-up steps and a cosine learning rate schedule, starting at 3e-5 with 4M batch size.
Detailed results can be found
here
