KAIST researchers introduced CluSTER, which clusters examples by similar training effects and distributes representative samples across GPUs in a balanced way. Across three coding and medical datasets and eight evaluations of 7-billion- and 13-billion-parameter models, the team reports comparable performance using about 31% of the data and up to 69.6% less training time. Tests centered on four GPUs, with benefits also observed on eight. Jointly optimizing selection and parallel allocation is the key contribution; the results do not guarantee the same cost reduction for every model or deployment.