Code and outputs for: Retrieval, Verification, and Latency Trade-offs for a Small Local Language Model: A Reproducible SciFact Pilot Study
Scripts, saved outputs and figures for a pilot study of Qwen2.5-3B-Instruct on SciFact: dense retrieval evaluation (300 queries), paired model-only versus RAG latency (100 queries), post-hoc verification, and a retrospective score-gated verification analysis. The SciFact corpus is not included and must be obtained from...