nvidia/Nemotron-SFT-Finance-Document-Grounded-Reasoning-v1

Dataset Description:

This dataset is designed to improve a model's understanding and reasoning over real financial filings (SEC 10-K, 10-Q, and 8-K filings from S&P 500 companies, 2017–2024). Questions and answers are synthetically generated by an LLM grounded in a given source document (document-grounded SDG).

Nemotron-SFT-Finance-Document-Grounded-Reasoning-v1 contains 63,056 document-grounded synthetic finance examples. The release retains only permitted limited context excerpts drawn from the bodies of Forms 10-K, 10-Q, and 8-K. The full SEC corpus, all exhibits, and any content bearing a third-party copyright notice are excluded, including consultant reports, presentations, fairness opinions, and analyst research.

This dataset is ready for commercial or non-commercial uses.

Dataset Owner(s):

NVIDIA Corporation

Dataset Creation Date:

Created on: July 27, 2026
Last Modified on: October 14, 2026

Versioning:

Current Version: 1.0 Previous Version(s): no previous version

License/Terms of Use:

Governing Terms: Use of this dataset is governed by OpenMDW-1.1. Each user is responsible for checking the content of the dataset and the applicable licenses and determining if suitable for the intended use.

Additional Information: This dataset includes excerpts from public filings obtained from the U.S. Securities and Exchange Commission EDGAR system (https://www.sec.gov/edgar). SEC filings are public records. Excerpts are provided for grounding only and remain the responsibility of their respective issuers.

Intended Usage:

This dataset is intended for LLM engineers and research teams developing and training large language models to improve document-grounded financial reasoning based on provided source context.

Dataset Characterization

Data Collection Method
Synthetic - LLM-generated financial question and answer pairs

The context excerpts are reproduced from public SEC EDGAR filings. No third-party license notice is required for SEC filing text.

Generation Models

• Qwen3.5-397B-A17B: question generation.
• DeepSeek-V4-Pro: answer generation.
• gpt-oss-120b: filtering and GenSelect.

Labeling Method
Synthetic - Model-generated question and answer pairs

Dataset Format

Modality: Text
Format: JSONL
Structure: Text + Metadata

Data file: final_result_sft_kept.jsonl, loaded as the train split.

Dataset Quantification

Records	Features	Storage
63,056	7	1.01 GB (1,010,575,788 bytes)

Features: Problem, Prompt, Source Context (included in messages[0].content), Answer (messages[1].content), Reasoning Content (messages[1].reasoning_content), and Question Type (metadata.question_type).

Reference(s):

• SEC EDGAR

Ethical Considerations:

Excerpts may contain names, titles, compensation and biographical information about company officers and directors as disclosed in public SEC filings. This is public business information disclosed under securities law. The content is not financial advice.

NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications. Developers should work with their internal developer teams to ensure this dataset meets requirements for the relevant industry and use case and addresses unforeseen product misuse.

Please report quality, risk, security vulnerabilities or NVIDIA AI Concerns here.