MetaMathQA/experiments/c3a/llama-3.2-3B-default/training_params.json

{
  "optimizer_kwargs": {
    "lr": 3e-1,
    "weight_decay": 1e-5
  }
}