NVIDIA has unveiled an iterative synthetic data generation pipeline engineered to resolve severe class imbalances in financial natural language processing datasets. By integrating NVIDIA NeMo Data Designer, NeMo Curator, and the Nemotron 3 Nano 30B model, the initiative produced 502,536 unique financial news headlines across 13 categories. The dataset directly enhances trading research, risk modeling, and regulatory surveillance by supplying balanced synthetic corpora for rare but critical financial events. Conventional financial data generation disproportionately captures routine market movements while neglecting niche occurrences like credit downgrades or supply chain disruptions. Initial single-pass attempts yielded a 65 percent duplicate rate, proving inefficient for large-scale production. To resolve this, developers implemented a closed-loop iterative workflow executed on a single eight-way NVIDIA B200 node over six days. Each cycle processes headlines through generation, rule-based quality filtering, global semantic deduplication, strategic few-shot selection, and dynamic category reweighting. The system leverages NeMo Data Designer to route prompts to a vLLM-served Nemotron model optimized for high-throughput synthesis. Generated outputs immediately undergo structural filtering to eliminate malformed or truncated text. The pipeline core mechanism relies on global semantic deduplication via NeMo Curator, which cross-references every new headline against the entire accumulated corpus using K-means clustering and a strict 0.90 cosine similarity threshold. To prevent computational bottlenecks as the dataset expands, the architecture utilizes 500 dynamic clusters, drastically reducing pairwise comparison overhead while maintaining semantic diversity. Post-deduplication, the system dynamically selects three representative few-shot examples per category to guide subsequent generations. Selection algorithms prioritize headlines positioned farthest from their cluster centroids while rejecting any candidate exhibiting greater than 80 percent semantic overlap with previously used prompts. This evolutionary prompting strategy actively directs the model toward novel content regions. Simultaneously, a distribution correction algorithm continuously adjusts category weights to counteract generative bias, ensuring underrepresented financial events achieve targeted representation. The workflow rapidly stabilized into an efficient production cycle. Early iterations retained approximately 35 percent of generated headlines, while later stages settled at 15 to 17 percent yield as cross-batch collisions increased. This reduced efficiency reflects corpus growth rather than quality degradation, as refined prompts consistently synthesized previously unexplored financial scenarios. The final dataset successfully balanced category distribution, with niche segments closely matching predefined targets. The resulting corpus provides a scalable, reproducible foundation for fine-tuning financial AI models. By demonstrating how global deduplication, dynamic weighting, and adaptive prompting overcome synthetic data homogeneity, the pipeline establishes a new operational standard for institutional research and enterprise AI development.