4 ms·

Disagree on Nvidia, most folks fine-tune model. Proof: there are about 20k models in huggingface derived from llama 2, all of them trained on Nvidia GPUs.

• eggdaft 2y ago
  Fine tuning can take a fraction of the resources required for training, so I think the original point stands.
  • nightski 2y ago
    Maybe in isolation when only considering a single fine tune. But if you look at it in aggregate I am not so sure.