nvidia/Nemotron-RL-Agentic-Finance-v1

Dataset Description:

This dataset contains 10,185 agentic reinforcement-learning tasks for financial reasoning over public SEC filings. Each example provides a financial question, a reference answer, the finance-agent prompt and tool definitions, and difficulty metadata collected from four target-model attempts. The questions cover Numerical Reasoning, Financial Modeling, Market Analysis, Adjustments, and Beat-or-Miss.

The tasks do not include supplied filing context. During training, the agent must use sec_filing_search, parse_html_page, and retrieve_information to find and reason over relevant SEC filings before submitting its answer.

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

Dataset Owner(s):

NVIDIA Corporation

Dataset Creation Date:

Created on: Aug 7, 2026
Last Modified on: Aug 7, 2026

Versioning:

v1.0

Previous Version(s): no previous version

Relationship to Previous Version(s): Not applicable - first release. This RL data was separated from finance_agentic_tool-calling_dataset_based_on_vals_v1_finance_benchmark v1.0, which was replaced by the SFT-only v2.0.

License/Terms of Use:

Use of this dataset is governed by OpenMDW-1.1.

Intended Usage:

This dataset is intended for researchers and developers post-training tool-using financial language models with reinforcement learning using NVIDIA NeMo RL and NeMo Gym. See the Nemotron-3-Ultra training guide for end-to-end training instructions.

Dataset Characterization

Data Collection Method
Synthetic. Questions were derived from public VALS-informed seeds and augmented using template-based and relational document-grounded SDG over public SEC filings. Multiple agent responses were generated and evaluated to construct a reference answer for each question. The target model then attempted each task four times in the finance-agent environment. Questions solved successfully in all four attempts were excluded, while consistently difficult and mixed-success questions were retained for RL.

The selected reference answer is stored in expected_answer and generation. The file does not contain the generated rollout trajectories themselves.

Labeling Method
Synthetic - Model-generated question and answer pairs

Dataset Format

Modality: Text
Format: JSONL
Structure: Text + Metadata

Important fields include:

• question and problem: Financial question
• expected_answer and generation: Reference answer
• prompt: Finance-agent instructions
• responses_create_params: Model input and available tool definitions
• agent_ref: Finance-agent configuration
• difficulty_profile: Four-attempt reward profile
• question_type: Financial reasoning category
• metadata.source_split: Train or validation assignment
• uuid: Unique record identifier

The context and reasoning_content fields are empty in all records.

Dataset Quantification

Split	Samples
Train	10,084
Validation	101
Total	10,185

Category	Samples
Numerical Reasoning	3,852
Market Analysis	2,113
Financial Modeling	1,680
Adjustments	1,638
Beat or Miss	902
Total	10,185

Record Count: 10,185 records (text/JSONL rows; no images, video or audio)
Feature Count: 15 top-level features per record
Features:

• question
• problem
• expected_answer
• generation
• prompt
• responses_create_params
• agent_ref
• difficulty_profile
• question_type
• current_date
• dataset
• metadata
• uuid
• context
• reasoning_content

Total Data Storage: 165.66 MB (165,660,351 bytes)

Reference(s):

1. NVIDIA NeMo RL
2. NVIDIA NeMo Gym
3. SEC EDGAR
4. Vals.ai Finance Agent Benchmark

Ethical Considerations:

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.