Domain-specific MultiVectorEncoder training on MIRIAD

training_contrastive.py trains a ColBERT-style late-interaction model for medical question -> passage retrieval on MIRIAD, as an example of adapting a multi-vector model to a specific domain with plain contrastive learning: no teacher scores or mined negatives are required, only (question, passage) pairs.

Because MIRIAD passages average ~940 tokens, the script doubles as a long-document late-interaction example: the max_length training argument caps tokenization during training while the saved model keeps the backbone's full context, and the evaluation calls run at full passage length without any manual memory tuning, because MaxSim scoring packs documents under an element budget on its own.

Reference results (nDCG@10, deduplicated passage corpus per split, single RTX 3090, ~9.5 hours):

Model	eval	test
multivector-gte-modernbert-base-miriad (this script)	0.9827	0.9826
splade-modernbert-base-miriad	0.8609	0.8626
pre-training baseline (warm-started backbone, random projection)	0.9176