We present a new, high-veracity chemical space dataset-bigQM7ω -with 12,880 molecules containing up to 7 heavy atoms, and highlight the key challenges in quantum machine learning modeling of electronic excitation spectra. We show excited state modeling with global structural representations to suffer from information overload resulting in diminished structure-property mapping. To improve the signal-to-noise ratio in the modeling, we use locally integrated spectral intensities and highlight a resolution-vs.-accuracy dilemma. Intensities derived from transition probabilities enable quantifying the prediction errors through probabilistic confidence scores. Upon changing the basis set at the target density functional theory level, > 75% confidence score is obtained only at the expense of the resolution, amounting to increasing uncertainties in peak positions. Compared to this, models with state-of-the-art structural representations trained only on < 10% of the data recover the full electronic spectra of the remaining molecules with higher confidence even for sub-nm wavelength resolutions.