wflying/nemotron-nano-rl-math-22k

Nemotron Nano RL Math 22K

Nemotron Nano RL Math 22K is a 22,056-example English mathematical reasoning dataset prepared for reinforcement learning with verifiable rewards (RLVR). It contains the DAPO-Math and Skywork math profiling components identified in the metadata of NVIDIA's Nemotron-3-Nano-RL-Training-Blend, represented in a compact prompt / label / metadata JSONL schema.

Each example contains a ready-to-use user prompt, a reference answer, and empirical pass-rate metadata from repeated model generations. The dataset contains problems and answer targets, not model responses or chain-of-thought traces.

Dataset summary

Property	Value
Examples	22,056
Split	train only
File	nemotron-nano-rl-math-22k.jsonl
Format	UTF-8 JSON Lines; one object per line
File size	24,183,051 bytes (about 23.1 MiB)
Unique metadata.question values	18,235
Prompt shape	One user message per example
Token count (metadata.num_tokens)	median 112; mean 125.3; 95th percentile 225; range 30–2,600
SHA-256	ab380cbae419b244a96ec818503eaaef40e8fbd0b01ed33569d5b8356f946cde

All 22,056 lines parse as JSON and contain the top-level fields prompt, label, and metadata.

Composition

Component	metadata.data_type	Examples	Share	Unique question texts
DAPO-Math profile	nemotron_math_dapo17k	10,224	46.35%	6,403
Skywork math profile (excluding Omni-Math)	nemotron_math_skywork_no_omni	11,832	53.65%	11,832
Total		22,056	100%	18,235

The corresponding internal dataset identifiers are nano_v3_sft_profiled_dapo17k and nano_v3_sft_profiled_skywork_no_omni.

The DAPO component contains 3,821 repeated occurrences of a metadata.question value; no question appears more than twice. The Skywork component has no repeated metadata.question values, and there is no exact question-text overlap between the two components. Repeated question text does not necessarily imply byte-identical records because prompts and profiling metadata can differ.

Upstream sources

• nvidia/Nemotron-3-Nano-RL-Training-Blend: upstream profiled training blend and dataset identifiers.
• BytedTsinghua-SIA/DAPO-Math-17k: source of the DAPO-Math problems.
• Skywork/Skywork-OR1-RL-Data: source of the Skywork math problems.

Data structure

Each line follows this structure:

{
  "prompt": [
    {
      "role": "user",
      "content": "Solve the problem ... Put your answer in \\boxed{}."
    }
  ],
  "label": "",
  "metadata": {
    "data_type": "nemotron_math_dapo17k",
    "num_tokens": 121,
    "agent_ref": {
      "type": "responses_api_agents",
      "name": "math_with_judge_simple_agent"
    },
    "pass_rate": 0.375,
    "pass_rate_total": 8,
    "pass_rate_passed": 3,
    "dataset": "nano_v3_sft_profiled_dapo17k",
    "question": "...",
    "expected_answer": "10"
  }
}

Fields

• prompt: A list containing exactly one user message. All prompts request a final answer using LaTeX \\boxed{} formatting.
• label: A compatibility field. It is an empty string in every record and must not be used as the reference answer.
• metadata.data_type: Reward/data routing identifier for the DAPO or Skywork component.
• metadata.num_tokens: Precomputed token count for the prompt.
• metadata.agent_ref: Verifier/agent configuration; all rows use responses_api_agents with math_with_judge_simple_agent.
• metadata.pass_rate: Fraction of profiled generations that passed the verifier.
• metadata.pass_rate_total: Number of profiled generations (8 or 16).
• metadata.pass_rate_passed: Number of generations that passed.
• metadata.dataset: Internal upstream component name.
• metadata.question: Source question text before or alongside prompt templating.
• metadata.expected_answer: Reference answer used for verification. This is the target field for training and evaluation.

For the DAPO component, prompt[0].content equals metadata.question. For the Skywork component, the prompt includes an answer-format instruction template around the source question, so consumers should use prompt as the canonical model input.

The file also contains 432 prompts with embedded Asymptote ([asy] ... [/asy]) diagram source.

Difficulty profiling

The pass-rate fields are internally consistent for every record:

pass_rate == pass_rate_passed / pass_rate_total

Component	Generation counts	Median pass rate	Mean pass rate	Range
DAPO-Math	8 for all 10,224 records	0.625	0.557	0–0.875
Skywork math	8 for 8,978 records; 16 for 2,854 records	0.250	0.363	0–0.9375
All records		0.500	0.453	0–0.9375

There are 5,312 zero-pass examples and no examples with a pass rate of 1.0. These values are empirical results for the profiling setup recorded by the upstream data pipeline; they should not be interpreted as model-independent measures of mathematical difficulty.

Loading the dataset

from datasets import load_dataset

dataset = load_dataset(
    "wflying/nemotron-nano-rl-math-22k",
    split="train",
)

example = dataset[0]
prompt = example["prompt"]
expected_answer = example["metadata"]["expected_answer"]
pass_rate = example["metadata"]["pass_rate"]

Do not train against label, which is empty by design. To remove the two records without a usable reference answer:

dataset = dataset.filter(
    lambda example: bool(example["metadata"]["expected_answer"])
)

For RLVR pipelines, metadata.agent_ref identifies the intended math-with-judge verifier family, while metadata.data_type can be used for source-specific routing or curriculum construction.

Data quality and validation

• 22,056 of 22,056 lines are valid JSON.
• Every record contains prompt, label, and metadata.
• Every prompt contains exactly one user message and a \\boxed{} answer-format instruction.
• Every profiling tuple satisfies pass_rate == pass_rate_passed / pass_rate_total.
• Every expected_answer value is a string; 22,054 are non-empty and 2 are empty.
• All top-level and metadata field types are consistent across the file.
• The released file's SHA-256 is included above for integrity checking.

The two empty-answer rows have a pass rate of 0 and should normally be filtered before answer-supervised or verifier-based training. Structural validation does not constitute a manual correctness review of every problem or expected answer.

Intended use

This dataset is intended for research and development involving:

• reinforcement learning with verifiable mathematical rewards;
• curriculum construction using empirical pass rates;
• evaluation of answer extraction and mathematical verification;
• controlled comparisons between DAPO-Math and Skywork math mixtures;
• mathematical reasoning post-training.

It should not be treated as a standalone benchmark because it contains only a training split and includes repeated question texts.

Limitations

• label is empty in every record; the reference answer is stored in metadata.expected_answer.
• Two records also have an empty metadata.expected_answer.
• The dataset has no validation or test split.
• The DAPO component contains 3,821 repeated question-text occurrences.
• Pass rates depend on the upstream model, sampling configuration, prompt template, and verifier; they are not intrinsic difficulty scores.
• The longest prompt has 2,600 tokens according to metadata.num_tokens.
• Asymptote diagrams may not render in all training or inspection tools.
• Source datasets may contain incorrect answers, malformed notation, ambiguous problems, contamination, or other source-specific issues.
• This compilation has not been independently audited for personal information, harmful content, or overlap with downstream evaluations.

Users should validate reward behavior, answer coverage, deduplication requirements, and train/evaluation separation for their application.

Licensing and attribution

This release uses the ODC Attribution License (odc-by) metadata of the upstream nvidia/Nemotron-3-Nano-RL-Training-Blend.

The DAPO source repository declares Apache-2.0. The Skywork source card does not currently expose a single license identifier in its Hub metadata. Users should review and comply with the NVIDIA blend license and the terms and attribution requirements of each upstream source. This section is informational and is not legal advice.

Citation

Please cite and attribute the upstream dataset releases used by the component(s) in your work:

• nvidia/Nemotron-3-Nano-RL-Training-Blend
• BytedTsinghua-SIA/DAPO-Math-17k
• Skywork/Skywork-OR1-RL-Data

The DAPO project paper is available as DAPO: An Open-Source LLM Reinforcement Learning System at Scale, and the Skywork-OR1 report is available at arXiv:2505.22312.