Base-model size effects persist after LoRA adaptation: the 7B model gained more on ROUGE-2 and ROUGE-L, yet 3B-r16 outperformed 7B-r4 on ROUGE-L and BERTScore-F1¶

Claim · All claims
Evidence · 1 study (2 entries) · causal r1–r2 · q2
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person · recheck by 2029: tool and model accuracy changes quickly

Subclaims¶

q2 i? After LoRA fine-tuning at r=16, the 7B model's advantage increased on ROUGE-2 and ROUGE-L while the ROUGE-1 gap narrowed in favor of the 3B model. → Shaowen Lu 2026 q2 i? A smaller model with higher-rank LoRA adaptation can outperform a larger model with lower-rank adaptation on some metrics. → Shaowen Lu 2026 (2)

Evidence¶

Shaowen Lu 2026¶

Shaowen Lu, Chengxu Liu, Ping Zhou and Tao Yang. (2026). LoRA Fine-Tuned Models for Control Systems Course Q&A: A Multidimensional Evaluation of Model Scale and Rank Effects. https://arxiv.org/abs/2609.13918 doi:10.48550/arxiv.2609.13918

q2 · i? · causal · r1

Direct comparison of base models and r=16 fine-tuned models on the 54-item test set (Section 5.5). The article reports the 3B model narrowed the ROUGE-1 gap while the 7B model gained more on ROUGE-2 and ROUGE-L under the same LoRA setting.

"The ROUGE-1 difference decreased from +0.0263 to +0.0148, indicating that the 3B model narrowed the lexical-overlap gap through LoRA adaptation. In contrast, the ROUGE-2 difference increased from +0.0199 to +0.0318, and the ROUGE-L difference increased from +0.0207 to +0.0317."

Shaowen Lu 2026 (2)¶

Shaowen Lu, Chengxu Liu, Ping Zhou and Tao Yang. (2026). LoRA Fine-Tuned Models for Control Systems Course Q&A: A Multidimensional Evaluation of Model Scale and Rank Effects. https://arxiv.org/abs/2609.13918

q2 · i? · causal · r2

Cross-size comparison on the current test set (Section 5.5). The article reports 3B-r16 achieved ROUGE-L 0.3776 versus 0.3615 for 7B-r4, and BERTScore-F1 0.8606 versus 0.8579, which the authors note is relevant for local deployment under limited resources.

"Its ROUGE-L score of 0.3776 exceeded the 0.3615 achieved by 7B-r4. Its BERTScore-F1 of 0.8606 likewise exceeded the 0.8579 of 7B-r4."

Discussion¶

Related Claims¶

• Bootstrap stability of LoRA gains is not uniform: the ROUGE-1 confidence interval crossed zero at 7B-r4, so that increase is a trend rather than a stable improvement — related
• LoRA fine-tuning on control systems course data increases lexical and semantic similarity between model responses and course reference answers at both 3B and 7B model sizes — related
• Increasing LoRA rank raises absolute ROUGE-L gains but produces diminishing returns in performance per million adapter parameters — related
• LoRA adaptation significantly increases closed-book accuracy on the PhysicsMate test set at all three Qwen3 scales — related
• Cluster membership showed a strong multivariate effect on standard text-similarity metrics (Pillai's trace = 0.85, F(12, 248) = 15.31, p < 0.001) — related
• Adaptation gains grow monotonically with model capacity under the same LoRA recipe, though three data points cannot establish a scaling law — related