Nemotron-SpecializedDomains-Finance is a large-scale synthetic financial question-and-answer dataset released by NVIDIA in 2025. The related paper isGenSelect: A Generative Approach to Best-of-N”, aimed at enhancing the performance of large language models on professional financial reasoning and document understanding tasks.

The dataset contains 326,698 high-quality question-answer pairs derived from SEC filings (10-K annual reports and 10-Q quarterly reports) of S&P 500 constituent companies between 2019 and 2024. The dataset employs a template-based synthetic data generation (SDG) method combined with the GenSelect filtering technique to ensure that all question-answer pairs are grounded in specific sections of authentic regulatory documents, covering domains such as corporate finance, risk factors, financial performance, governance, and compliance. The data is pre-formatted in a conversational structure (system/user/assistant) and designed specifically for supervised fine-tuning (SFT), making it suitable for applications such as developing financial assistants, compliance analysis, and investment research.

Dataset Composition

The dataset includes only the train split, with each sample containing the following fields:

• messages: An array of role-based conversational structures, including system (empty in this dataset), user (containing financial document context and questions), and assistant (containing reasoning steps and answers) messages.
• metadata: Contains a unique identifier (UUID) and the identifier of the synthetic data generation model (e.g., "openai/gpt-oss-120b").