amityco/lbm-v3-4-r2

LBM v3.4-r2 — SEA consumer digital-twin (coupon-ranking + price-coherence specialist)

Data & usage notice. This model was finetuned on a proprietary loyalty-card transaction dataset from a Southeast-Asian grocery retailer (not distributed; the training-data repos in this org are private). Like any LLM finetune, it may have memorized fragments of its training data — treat generated personas, baskets and behaviors as SYNTHETIC simulations, never as records of real individuals, and do not attempt to extract or re-identify training examples. Proprietary license: research/evaluation use within the terms set by the publishing org.

Single stage-2 SFT from v3.2's own base on amityco/lbm-sft-v34r2 (v3.2's mix byte-identical

• 104k think-format SG lanes) — identical recipe to v3.2, so the two models differ only in data. Reproduce: repro/train/train_v34r2.sh. Evals: repro/eval/*.sh (protocols pinned).

Summary results (2026-07-28 symmetric battery)

Pre-registered gates: G1 persona ov@5 0.323 ✅ · G2 price-ladder monotonicity 99.5% ✅ (the promotion blocker — its predecessor scored 0.0%) · G3 Lazada 0.670 ✅ · G4 HES income gradient MAE 2.12 ❌ (bar 1.6) · G5 WVS-TH TVD 39.4 vs bar ≤29 ❌ (4-shard protocol rerun 2026-07-28).

Instrument	v3.4-r2	v3.2	notes
§7 coupon (proprietary TH retailer, 2,000 real events) AUC	0.721	0.697	best of any SG arm
Lazada 12,480 AUC	0.670	0.661
IDN PSM strict-monotone / PME err	99.5% / 9.2%	1.0% / 21.8%	the r2 fix
SG persona ov@5 / rating MAE	0.323 / 1.419	0.084 / 3.056
§7 mean P(redeem) (real .199)	0.815 ⚠	0.759	most over-confident — recalibrate before any volume forecast
Brand's WTP r / bias / MAPE / T2B	0.878 / −29.2% / 32.8% / 0.100	0.990 / −10.2% / 19.2% / 0.547	S$-register anchoring suspected; do not use for THB client pricing
HES income / age gradient MAE	2.12 / 1.89	2.03 / 1.75	G4 fail
HBA TVD / top-match	39.7 / 63%	36.1 / 63%	terse ×8-seed

Use v3.4-r2 for: top-K coupon/offer targeting (ranking), coherent WTP ladder elicitation, persona-conditioned interest tasks. Do not use for: volume forecasts without an isotonic recalibration layer, or THB client pricing surveys. Full analysis: docs/reports/v34r2_battery/REPORT.md.