---
pretty_name: Nemotron-Math-Proofs-v3-RL
language:
- en
license:
- cc-by-4.0
task_categories:
- text-generation
tags:
- math
- proofs
- mathematical-reasoning
- text
- synthetic
- automated
- post-training
- reinforcement-learning
- rlvr
- Nemotron_3_Ultra
size_categories:
- 1K<n<10K
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train.jsonl
---

# Nemotron-Math-Proofs-v3-RL

## Dataset Description:
Nemotron-Math-Proofs-v3-RL is a long-form mathematical reasoning dataset for reinforcement learning. The release contains 9,597 proof-generation prompts.

The dataset uses [NeMo Gym](https://github.com/NVIDIA-NeMo/gym)-compatible, single-turn user prompts derived from hard proof problems in the [AoPS](https://artofproblemsolving.com/) subset of [nvidia/Nemotron-Math-Proofs-v1](https://huggingface.co/datasets/nvidia/Nemotron-Math-Proofs-v1). The `train` split asks the policy to produce a rigorous solution and self-evaluation. Policy responses and realized rewards are generated during training and are not stored in the file.

Full details about dataset construction can be found in our technical report [An Open Recipe for IMO Gold: Training Nemotron for Olympiad Mathematics](https://arxiv.org/abs/2609.10712).

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

## Dataset Owner(s):
NVIDIA Corporation

## Dataset Creation Date:
Created on: 07/01/2026

## Versioning:
Nemotron-Math-Proofs-v3-RL

Previous Version(s):
* [nvidia/Nemotron-Math-Proofs-v1](https://huggingface.co/datasets/nvidia/Nemotron-Math-Proofs-v1) : Previous proof dataset version used as the source of problems for this release.
* [nvidia/Nemotron-Math-Proofs-v2](https://huggingface.co/datasets/nvidia/Nemotron-Math-Proofs-v2) : Previous release version.
* [nvidia/Nemotron-Cascade-2-SFT-Data](https://huggingface.co/datasets/nvidia/Nemotron-Cascade-2-SFT-Data) : Contains a natural language proof subset from the same source problems.

Relationship to Previous Version(s): This release adds reinforcement-learning data for proof generation, based on the same family of hard proof problems used for Nemotron-Math-Proofs-v1.

## License/Terms of Use:
This dataset is governed by the [Creative Commons Attribution 4.0 International License (CC BY 4.0)](https://creativecommons.org/licenses/by/4.0/).

## Intended Usage:
This dataset is intended for:

* Reinforcement learning for structured mathematical reasoning and proof generation using online verifier feedback.
* Training LLMs to produce proof self-evaluations and identify gaps in mathematical arguments.
* Research on proof generation and reinforcement learning with verifiable rewards.

## Dataset Characterization

### Dataset Composition and Generation

#### Problem Sourcing
Problems are derived from [nvidia/Nemotron-Math-Proofs-v1](https://huggingface.co/datasets/nvidia/Nemotron-Math-Proofs-v1), which collected proof-based problems from the [AoPS](https://artofproblemsolving.com/) community. The released records identify their source as `AoPS`.

#### Reinforcement-Learning Data Preparation
The file is formatted for NeMo Gym and routes examples through `agent_ref`:

* **Proof generation:** Each raw problem is inserted into a prover prompt and routed to `proof_simple_agent`. The expected policy response contains a solution, a self-evaluation, and a boxed score. During training, a proof-verification judge can score the generated proof and optionally assess the self-evaluation.

#### Dataset Fields
The `train` split contains the following fields:

* `agent_ref`: NeMo Gym agent routing information.
* `responses_create_params`: A single user message containing the formatted policy prompt.
* `problem`: The natural-language mathematical problem.
* `uuid`, `license`, `source`, and `dataset`: Record identifier and release provenance fields. All released records use `cc-by-4.0`, `AoPS`, and `Nemotron-Math-Proofs-v3`, respectively.
* `metadata`: An empty list in the released file.

#### Data Collection Method
* Hybrid: Manually-Collected, Synthetic, Automated

#### Labeling Method
* Hybrid: Manually-Labeled, Synthetic, Automated

## Dataset Format
* **Modality:** Text
* **Format:** JSONL
* **Structure:** A single `train` split containing proof-generation prompts for reinforcement learning.

## Dataset Quantification
Split Samples:

| Split | File | Samples | Unique Problems | Disk Size |
|---|---|---:|---:|---:|
| `train` | `train.jsonl` | 9,597 | 9,597 | 50,430,007 bytes |

The reinforcement-learning file does not contain generated-token counts because it stores policy inputs rather than completed policy rollouts.

## Reference(s):
* [An Open Recipe for IMO Gold: Training Nemotron for Olympiad Mathematics](https://arxiv.org/abs/2609.10712) (technical report)
* [DeepSeek-V4-Pro](https://huggingface.co/deepseek-ai/DeepSeek-V4-Pro)
* [nvidia/Nemotron-Math-Proofs-v1](https://huggingface.co/datasets/nvidia/Nemotron-Math-Proofs-v1)
* [nvidia/Nemotron-Math-Proofs-v2](https://huggingface.co/datasets/nvidia/Nemotron-Math-Proofs-v2)
* [DeepSeekMath-V2: Towards Self-Verifiable Mathematical Reasoning](https://arxiv.org/abs/2511.22570)
* [Nemotron-Math paper](https://arxiv.org/abs/2512.15489)
* [NeMo-Skills](https://github.com/NVIDIA/NeMo-Skills)
* [Nemotron-MathProofs-v1 documentation](https://nvidia-nemo.github.io/Skills/releases/nemotronmathproofs/)

## 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](https://www.nvidia.com/en-us/support/submit-security-vulnerability/).