Skip to content

Tree Model Formats

LightGBM and XGBoost models are serialized as columnar tensors by default: split features, thresholds, child pointers, and leaf values are stored as typed numpy arrays in safetensors, with only small scalar config (objective, feature names, etc.) in JSON.

This reduces file size by 30-85% compared to the native format and makes the JSON human-readable.

Using the native format

To use the library's own format instead:

skeights.save(model, "model.safetensors", "model.json", format="native")
  • LightGBM native: stores the booster model string (LightGBM's text format) in JSON.
  • XGBoost native: stores the booster model as XGBoost's own JSON dict.

Loading

The loader automatically detects whether a model was saved in columnar or native format and handles both transparently. You do not need to specify the format when loading.