ARCHIVES

Small language models could support several product-understanding operations within a retailer’s infrastructure, but evidence of task accuracy is easily con-founded with evidence of generalization or serving efficiency. We study QLoRA adaptation of Qwen3-1.7B for query–product relevance, listing identity, functional relation, technical compatibility, and brand/color extraction. We reconstruct the experiments from committed data, checkpoint histories, training logs, prediction records, and serving traces, and release an offline audit that regenerates the quanti-tative analysis. On a common 200-listing extraction sample, the dedicated adapter attains brand F1 of 1.000 and color F1 of 0.955 under the recorded permissive matching rule; the latter exceeds the strongest tested API comparator by 1.38 per-centage points. This is a descriptive result: two evaluation texts also occur in training, and missing paired outputs prevent a significance claim. Matching re-sults expose a different problem. Adapter and API relevance scores use different samples; a single-token relevance adapter reaches 0.5115 accuracy but only 0.436 macro-F1; and 25 of 28 functional-relation test rows reuse training scenarios. A checkpoint sweep further exposes that test partition to model selection. We for-malize the information required for comparable evaluations, demonstrate scorer counterexamples, and bound the paired correctness uncertainty recoverable from aggregate extraction counts. Historical load tests contain no completed inference requests. A separate RunPod follow-up adds paired development evaluations un-der two prompt interfaces, a text-extraction comparator and bounded HTTP serv-ing measurements; these do not establish final-test generalization or production cost advantage. The evidence supports promising narrow extraction behavior and a reproducible account of evaluation failure modes, rather than universal superior-ity or a production-ready horizontal model.