How large must the validation set be when selecting a merge coefficient? We take an 11-point linear-interpolation grid between Qwen2.5-0.5B and its Instruct variant, scored by the macro average of four multiple-choice benchmarks. We measure selection regret (test accuracy of the best grid point minus that of the selected point) on random validation subsets of n∈30,60,100,200 items per task, 1000 draws each, from a pool of about 300 items per task. The mean regret is 0.50, 0.53, 0.51 and 0.50 pp. We detect no trend, and the pre-registered hypothesis that regret falls with n is not supported. Selection does not change the expected regret, although the probability of zero regret roughly doubles (0.10 to 0.20). The probability of beating both originals (0.31–0.37) is of the same order as that of a random pick (0.36), whose regret is 0.46 pp. The test accuracies of all 11 points lie within 1.06 pp; we conjecture that this flatness sets the regret. One model pair, one grid, one pool.
Download the PDF · Figshare record and reproducibility package