An NVIDIA technical blog describes an iterative pipeline for generating synthetic financial news headlines to address data limitations in fine-tuning LLMs for financial NLP. Real-world data overrepresents earnings and stock movements while underrepresenting rarer events like credit-rating changes, product approvals, and labor issues. The workflow combines NVIDIA NeMo Data Designer, NeMo Curator, and Nemotron models to generate 500,000 unique headlines across 12 topics plus an “Other” category. A naive run of 50,000 headlines saw 65% removed as near-duplicates. The iterative pipeline generates, filters, deduplicates globally, selects distinctive few-shot examples, corrects category weights, and repeats. The full run produced 502,536 unique headlines across 13 categories in 82 iterations, using approximately 6 days of compute on a single 8-way NVIDIA B200 node. Pipeline parameters included 35K headlines per batch (50K in early iterations), a 90% cosine-similarity deduplication threshold, 500 K-means clusters, 3 few-shot examples per category, and an 80% cross-iteration similarity cutoff for example selection.