In-the-wild eye tracking datasets to train gaze vector estimation methods are noisy and unreliable, due both to limited image quality and to weak guarantees on label accuracy. At the same time, simply increasing dataset size does not guarantee better results, especially when diversity, balance and label reliability are not taken into account. In this work, we adopt a data-centric perspective and introduce an innovative data-pruning pipeline for robust gaze vector estimation. Our methodology leverages vision-language model embeddings and their zero-shot classification capabilities, together with an early-learning prediction-error analysis, to assign task-aware quality scores to individual samples. Moreover, we incorporate a region-weighted loss conditioned on gaze direction to better handle the uneven difficulty across the gaze space and to improve learning in under-represented regions. Finally, we validate our approach on a large relabeled subset of the GazeCapture dataset by training a hybrid gaze-estimation model, and show that a curated subset of 84K samples yields an accuracy improvement over a model trained on the full 2.45M-sample dataset (https://dx.doi.org/10.21227/c9ap-zm59 doi:10.21227/c9ap-zm59).