Running mean bias and mean dispersion as a function of labels for the train (blue), test (orange), and observed (green) sets calculated in bin sizes: 250 K for T_eff, 0.3 dex for log(g), [Fe/H], and A(Li). The curves are representative of the real accuracy and precision of our CNN predictions. Bias = mean(CNN-iDR6) and σ = std(CNN-iDR6) for each bin. On the right column we present the distribution of the train, test and observed sets in logarithmic y-axis. The observed sample is selected within the training set, with S/N > 20 pix⁻¹ and no GES flags; for A(Li), we selected only stars with Li measurements, instead of those with upper limit Li estimates.