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

{
  "optimizer_kwargs": {
    "lr": 1e-3
  }
}