{
  "dataset_split": {
    "train": {
      "file": "train.extxyz",
      "description": "Training set. 90% of this file is used for actual training; the remaining 10% is automatically held out as the validation set by the MACE run_train script."
    },
    "validation": {
      "file": "train.extxyz",
      "fraction": 0.10,
      "source": "automatic",
      "description": "Validation split is not a separate file. It is carved out automatically from train.extxyz by MACE at training time, controlled by the --valid_fraction=0.10 argument in the training bash script."
    },
    "test": {
      "file": "test.extxyz",
      "source": "manual",
      "description": "Independent test set, collected and set aside separately. Not derived from train.extxyz"
    }
  },
  "notes": [
    "valid_fraction=0.10 in fine_tune.sh determines the validation split size (10% of train.extxyz).",
    "test.extxyz is a standalone holdout set, kept fully separate from the train/validation split."
  ]
}