ARCHIVES

Small language models offer a practical way to convert shopping queries into ex-ecutable constraints, but partial extraction accuracy can obscure whether the com-plete request is represented correctly. We study CommerceCore, a Qwen3-1.7B model adapted with QLoRA on 2,130 examples spanning entity segmentation, query constraints, and catalog attributes. The recorded run takes 75.2 minutes at an estimated GPU cost of $0.90. Archived served-model predictions reproduce mean constraint-set F1 of 0.683 on 20 development queries, with seven exact-set matches; six harder diagnostic queries yield F1 of 0.400 and no exact-set matches. A separate aggregate records entity-set F1 of 0.498 on 993 QueryNER examples. We connect these results to the complete training-loss trace, eight checkpoint eval-uations, and an executable audit of data and evaluation integrity. The audit iden-tifies eight groups of inconsistent synthetic targets, test examples used for check-point selection, and an API baseline prompt covering only two of 17 entity labels. These conditions limit comparative and generalization claims. The contribution is a reproducible empirical account of low-budget adaptation and the gap between partial extraction scores and faithful constraint sets, with public code, model arti-facts, and result-level provenance.