training/DPO/__init__.py

"""
DPO (Direct Preference Optimization) module for LKAlert alert-timing alignment.

Aligns HazardHead to prefer timely alerts (TTA ∈ [1.5, 5.0]s) over
too-early, too-late, or false-alarm predictions.

Stage flow:
  1. make_dpo_pairs.py  — build preference pair manifests from SFT manifests
  2. trainer.py         — DPO fine-tune HazardHead on top of frozen SFT model
"""

from .dataset import DPODataset, dpo_collate_fn
from .trainer import DPOModel, DPOTrainer, compute_dpo_loss

__all__ = [
    "DPODataset",
    "dpo_collate_fn",
    "DPOModel",
    "DPOTrainer",
    "compute_dpo_loss",
]