Rank Is Not the Bottleneck: Data Scale, Not Adapter Capacity, Drives Faithful Summarization in LoRA-Tuned LLMs

In this study, however, we find that increasing the LoRA rank has very little effect on summarization performance. In contrast, increasing the amount of fine-tuning data produces substantial improvements across nearly all of the evaluated metrics, particularly those related to factual grounding. Our results show that doubling the LoRA rank provides almost no meaningful improvement. With the full training set, (r=16) achieves a ROUGE-1 score of 37.82, compared with 37.70 for (r=8). With only (10^3) training examples, the ordering actually reverses: (r=8) achieves 28.61, while (r=16) reaches 28.17. These small and inconsistent differences suggest that the effect of increasing adapter capacity is largely negligible under our experimental conditions. The amount of training data, on the other hand, has a much stronger effect, especially on factual faithfulness. MiniCheck claim support increases only slightly from 0.496 for the zero-shot model to 0.510 when fine-tuned on (10^3) examples. When the full training set is used, however, the score rises substantially to 0.680. This suggests that factual grounding in this setting is driven much more by the amount of fine-tuning data than by the number of LoRA parameters. We also observe an important difference between semantic similarity and factual grounding. BERTScore F1 changes non-monotonically across the same conditions, moving from 0.510 to 0.626 and then to 0.600, while MiniCheck improves consistently as the amount of training data increases. This indicates that higher embedding-based similarity does not necessarily correspond to better claim-level factual support. Finally, the benefits of fine-tuning do not transfer equally across domains. The resulting adapter provides a modest improvement on XSum, but performance drops on the dialogue-based SAMSum dataset. The fine-tuned model achieves a score of 20.53 on SAMSum compared with 23.27 for the zero-shot baseline. In this case, fine-tuning actually makes the model perform worse than the original base model. Overall, these results suggest a simple but important conclusion: for LoRA-based abstractive summarization, increasing the amount of high-quality fine-tuning data matters far more than simply increasing the adapter rank.