onf-c2/src/onf/graph/train/__init__.py

"""Training the retrieval head: query construction, loss, held-out metrics, and the loop.

loop.py is the entry point (train_graph), data.py builds the queries, drift.py generates the
corruption curriculum, loss.py and metrics.py are the training step and the held-out evaluation,
types.py the parameter objects threaded through them.

Deliberately docstring-only: re-exporting train_graph would make importing the
TYPE_CHECKING-only types module pull in torch, onf.field and the whole trainer. Use
from onf.graph.train.loop import train_graph.
"""