amityco/lbm-v3.6-dpo

LBM v3.6 DPO

This is a merged, vLLM-ready LBM v3.6 checkpoint specialized for the Thai market. It starts from amityco/lbm-v3.6 and applies same-decision DPO on fresh, externally judged Thai customer rollouts.

Output contract

First-person Rationale -> Acceptance Probability -> Final Decision: ACCEPT/REJECT

Frozen 2,000-row Thai diagnostic

Task	ROC AUC
Coupon response	0.6811
Promotion response	0.6583
New-product adoption	0.7271

Gate pass: True.

Scope and limitations

• Behavior events use customer-held-out Lotus-derived ground truth and opportunity proxies.
• The small diagnostic is outcome-balanced and is not a population calibration estimate.
• HBA/WTP remain evaluation-only. v3.6 post-training does not yet dominate older LBM checkpoints on every survey metric, so run the complete frozen Thai benchmark before deployment.
• Do not interpret predicted acceptance as causal coupon lift.