NVIDIAVerified interview challenge
Curate the next foundation-model training mixture
Choose which data work to fund before a foundation-model training run commits its token mixture.
Unlock with Full AccessThe challenge
A base-model training run will lock its data mixture in ten weeks. The team can improve filtering and human review, license additional data, generate synthetic examples or fund small comparison runs, but the budget cannot cover every option.
Use the supplied source-quality and capability evidence to allocate the investment and propose the training mixture. Your plan must fit processing capacity and review deadlines while showing how the team will distinguish a useful curation change from one that only looks promising on paper.
What you'll practise
Portfolio decisions
Allocate funding across competing data-quality, coverage and evaluation investments.
Quality and provenance
Account for duplication, rights, privacy and contamination in the proposed mixture.
Learning before commitment
Design a feasible sequence of proxy experiments and decisions before the training run.
Connect the data choices to model gaps, budget and evidence.