How can we train a LLM from scractch in R with the R package torch?

Training an LLM from scratch in R using PyTorch involves defining a model, preparing a large tokenized text dataset, and running a training loop with cross entropy loss. For example, create embeddings, positional encoding, blocks, and an output layer, then optimize with Adam. True LLM training requires massive data, GPUs, and careful hyperparameter tuning; small models can be experimented with locally.

library(torch)
=========================================================
0) General config
=========================================================
set.seed(123)
cfg