Dataset and Backbone Effects on Accuracy–Efficiency Asymmetry in Hybrid Prefix–Prompt Tuning for Sentiment Analysis

Parameter-efficient fine-tuning (PEFT) reduces adaptation costs, but the predictive trade-offs may vary across datasets and encoder backbones. This study evaluated Hybrid Prefix–Prompt tuning for sentiment classification using BERT-base-uncased on 3000-example training and 500-example held-out subsets of SST-2 and IMDb. Five training seeds supported comparisons with full fine-tuning, LoRA, adapters, BitFit, and IA3; independently tuned Prefix-only and Prompt-only variants and targeted diagnostics were analyzed separately. Hybrid trained 311,810 parameters (0.284% of the model) and reduced peak training GPU memory by 46.8% relative to full fine-tuning, but inference was slower. The mean Hybrid accuracy was 86.44% on SST-2 and 80.80% on IMDb, versus 88.48% and 86.56% for full fine-tuning. Seed-level retention differed by 4.36 percentage points between datasets (95% CI: 1.34–7.38). Longer-context IMDb tests improved accuracy without eliminating the Hybrid deficit. RoBERTa diagnostics suggested that the reversed retention ordering was sensitive to learning rate selection; a tuned single-seed check retained the reversed direction but did not establish a causal backbone interaction. A retrospective restricted validation comparison favored Hybrid, whereas the held-out ablations did not show consistent synergy. The results underscore dataset- and optimization-sensitive accuracy–efficiency trade-offs rather than universal PEFT superiority.