howtospark
Training

Knowledge distillation

DistillationReference

Train a smaller student on a teacher's outputs; fused with quant as QAD to recover FP4 accuracy.

Relieves
CapacityBandwidthCompute
Targets
Weights
Format
Smaller student model
Granularity
Whole-model
Lifecycle
QAT / QAD · needs gradients
Calibration
Full corpus
Compression
Arbitrary (student size)
Quality
High with a good teacher; expensive to produce
Hardware
Training-grade compute
Runtimes
Transformers

A worked Spark recipe for this method hasn't been written yet — it lives here as a reference point in the ontology.