It's impossible. Meta itself cannot reproduce the model. Because training is randomized and that info is lost. First samples a coming at random. Second there are often drop-out layers, they generate random pattern which exists only on GPU during training for the duration of a single sample. Nobody saves them, it would take much more than training data. If someone tries to re-train the patterns will be different, which results in different weight and divergence from the beginning. Model will converge to something completely different. With close behavior if training was stable. LLMs are stable.
So, no way to reproduce the model. This requirement for 'open source' is absurd. It cannot be reliably done even for small models due to GPU internal randomness. Only the smallest trained on CPU in single thread. Only academia will be interested.
Interesting. LLAMA is trained using 16K GPUs so it would have taken around a quarter for them. An hour of GPU use costs $2-$3 so training a custom solution using LLAMA should be atleast $15K to $1M. I am trying to get started with this thing. A few guys suggested 2 GPUs were a good start but I think that would only be good for 10K training samples.
