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.
