amityco/lbm-v3-2

LBM v3.2 — SEA consumer digital-twin (flagship, client-instrument 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.

Lineage: Qwen3.5-27B → CPT (SEA consumer voice) → stage-1 SFT (public-source personas+reviews) → stage-2 SFT on amityco/lbm-sft-v32 (private) (LoRA, LR 1e-4, 1 epoch, seq 2048, 8×H200). Reproduce: repro/train/train_v32.sh. Evals: repro/eval/*.sh (protocols pinned).

Summary results (2026-07-28 symmetric battery; best-of-three-arms in bold)

Instrument	v3.2	v3.4-r2	notes
Brand's WTP: structure r / level bias / MAPE / T2B	0.990 / −10.2% / 19.2% / 0.547	0.878 / −29.2% / 32.8% / 0.100	8-seed protocol, anchor = published row
HBA panel TVD / top-match	36.1 / 63%	39.7 / 63%	terse ×8-seed protocol
§7 coupon (proprietary TH retailer, 2,000 real events): AUC / Brier / meanP (real .199)	0.697 / 0.469 / 0.759	0.721 / 0.534 / 0.815	exact repro of published 0.6973
Lazada 12,480 AUC	0.661	0.670
IDN PSM strict-monotone / PME err	1.0% / 21.8%	99.5% / 9.2%	v3.2 has the ladder-collapse issue
SG persona ov@5	0.084	0.323
HES income / age gradient MAE	2.03 / 1.75	2.12 / 1.89	v3.4-full leads (1.42/1.62)
Bangkok Likert r / MAE	0.393 / 0.420	0.363 / 0.489	8-seed
use-cases: churn / deal / purchase AUC	0.780 / 0.624 / 0.563	0.768 / 0.638 / 0.547

Use v3.2 for: client pricing surveys (Brand's-style), survey-panel replication, churn. Do not use for: persona-conditioned interest prediction (ov@5 0.084) or open-ended VW price ladders (1% monotone). Full analysis: docs/reports/v34r2_battery/REPORT.md.