The Atmospheric Machine Learning Experiment Competition (AMLEC), organized within the EU ELIAS project, provided a benchmark for evaluating machine learning approaches to emulating atmospheric radiative transfer models. Participants were tasked with predicting spectral data across two scenarios involving different input variables and spectral configuration: (A) atmospheric correction of hyperspectral satellite data, and (B) CO \(_2\) concentration retrieval. Several training datasets, covering realistic input ranges with 500 to 10,000 samples, were used. Testing included interpolation and extrapolation to out-of-range conditions. Eight models were submitted, spanning neural networks (3) and Gaussian processes (5) with various configurations. Results showed that Gaussian process approaches achieved the lowest errors, indicating their suitability while highlighting the challenge of training complex neural network approaches with scarce data. Beyond scientific insights, AMLEC fostered community engagement, reproducible workflows, and open data sharing. This paper summarizes the competition design, datasets, evaluation metrics, and key findings, providing lessons for future benchmarks at the intersection of atmospheric science and machine learning.