amityco/lbm-v3-4-r2
The amityco/lbm-v3-4-r2 is a 27 billion parameter language model developed by amityco, specifically fine-tuned for consumer digital-twin simulations in Southeast Asia. It excels at coupon ranking, offer targeting, and maintaining price coherence, demonstrating significant improvements in price-ladder monotonicity. This model is optimized for simulating consumer behavior, particularly for loyalty-card transaction analysis and persona-conditioned interest tasks.
Loading preview...
LBM v3.4-r2: Southeast Asian Consumer Digital-Twin Specialist
The amityco/lbm-v3-4-r2 is a 27 billion parameter model from amityco, fine-tuned on a proprietary loyalty-card transaction dataset from a Southeast Asian grocery retailer. It represents a single stage-2 SFT from its v3.2 base, with key differences stemming from updated training data. This model is designed to simulate synthetic consumer personas, baskets, and behaviors, rather than reflecting real individuals.
Key Capabilities & Performance
This iteration shows notable improvements over its predecessor, v3.2, particularly in maintaining price coherence and coupon ranking. Key performance highlights include:
- Coupon/Offer Targeting: Achieved an AUC of 0.721 on proprietary TH retailer coupon events, outperforming
v3.2. - Price-Ladder Monotonicity: Demonstrated a significant improvement with 99.5% strict-monotone performance, a critical fix over
v3.2's 1.0%. - Persona Simulation: Showed enhanced persona overlap (
ov@5) at 0.323 and reduced rating MAE compared tov3.2.
Recommended Use Cases
Use v3.4-r2 for:
- Top-K coupon/offer targeting and ranking.
- Eliciting coherent willingness-to-pay (WTP) ladders.
- Persona-conditioned interest tasks and consumer behavior simulations.
Limitations
- Volume Forecasts: Requires an isotonic recalibration layer before use for volume forecasting due to over-confident
P(redeem)predictions. - Pricing Surveys: Not recommended for THB client pricing surveys, as S$-register anchoring is suspected, affecting WTP metrics.