# Fine-tuning and evaluation

The final decoder uses metric local refinement (`metric_local_refinement`), with 128 local channels,
two blocks, and at most 16 neighbors. Both schedules run up to 120 epochs,
including ten frozen-encoder epochs, followed by cosine fine-tuning.
The QL1 and QL2 trainers preserve their respective research protocols.

Copy `configs/ql*_segmentation.yaml`, edit data paths and choose a new
output directory. Download the paired encoder artifact before fine-tuning.
A complete segmentation checkpoint is sufficient for inference.

```bash
python scripts/finetune_segmentation.py --config configs/ql2_segmentation.yaml
python scripts/evaluate_segmentation.py --config configs/ql2_segmentation_eval.yaml
```

The evaluation configs load released best segmentation checkpoints and use FP32.
When evaluating a new run, update `model.segmentation_checkpoint`.
QL1 includes the final common-support evaluation seed. mIoU uses five
unthresholded semantic classes with unmapped labels ignored. Confidence
thresholding belongs to exported products. QL2 evaluation is on the thinned
QL1 proxy. Crop visualization LAS files cannot reconstruct an original LAS
header or CRS from NumPy input.

Record any change to support sampling, normalization, labels, physical radii,
checkpoint, or split as a new experiment. Matching architecture alone does
not reproduce scores: original splits and support selection also matter.