data/sft/processed/train_metadata.json

{
  "num_examples": 200000,
  "seq_len": 2048,
  "input_ids_path": "train_input_ids.bin",
  "labels_path": "train_labels.bin"
}