| \begin{table}[t] | |
| \centering\small | |
| \caption{Dataset-size ablation. Accuracy (\%) of the restored model on MATH-500 under the primary configuration (\texttt{qwen25-7b} / GPTQ w4) as the silver-bullet dataset size $N$ varies. 95\% bootstrap CIs in brackets. $\Delta$ is restored $-$ quantized baseline. The $N\!=\!500$ row of this table (75.4\%) does not match the corresponding cell in Table~\ref{tab:main} (74.8\%) to the last decimal because the ablation is a separate QLoRA training run from the main matrix; the $\sim$0.6\,pp gap is consistent with the run-to-run variance documented in \S\ref{sec:setup} and the across-seed band in \S\ref{sec:multi-seed}.} | |
| \label{tab:ablation} | |
| \begin{tabular}{@{}rccc@{}} | |
| \toprule | |
| $N$ & Acc (\%) & 95\% CI & $\Delta$ vs.\ baseline (pp) \\ | |
| \midrule | |
| 50 & 74.6 & [70.6, 78.2] & +5.2 \\ | |
| 100 & 73.8 & [70.0, 77.6] & +4.4 \\ | |
| 250 & 75.6 & [71.8, 79.2] & +6.2 \\ | |
| 500 & 75.4 & [71.4, 79.2] & +6.0 \\ | |
| \bottomrule | |
| \end{tabular} | |
| \end{table} | |