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](https://huggingface.co/datasets/nvidia/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](https://huggingface.co/datasets/nvidia/Nemotron-3-Nano-RL-Training-Blend): upstream profiled training blend and dataset identifiers. - [BytedTsinghua-SIA/DAPO-Math-17k](https://huggingface.co/datasets/BytedTsinghua-SIA/DAPO-Math-17k): source of the DAPO-Math problems. - [Skywork/Skywork-OR1-RL-Data](https://huggingface.co/datasets/Skywork/Skywork-OR1-RL-Data): source of the Skywork math problems. ## Data structure Each line follows this structure: ```json { "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: ```text 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 ```python from modelscope.msdatasets import MsDataset dataset = MsDataset.load( "wflying000/nemotron-nano-rl-math-22k", subset_name="default", 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: ```python 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](https://huggingface.co/datasets/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](https://huggingface.co/datasets/nvidia/Nemotron-3-Nano-RL-Training-Blend) - [BytedTsinghua-SIA/DAPO-Math-17k](https://huggingface.co/datasets/BytedTsinghua-SIA/DAPO-Math-17k) - [Skywork/Skywork-OR1-RL-Data](https://huggingface.co/datasets/Skywork/Skywork-OR1-RL-Data) The DAPO project paper is available as [DAPO: An Open-Source LLM Reinforcement Learning System at Scale](https://arxiv.org/abs/2503.14476), and the Skywork-OR1 report is available at [arXiv:2505.22312](https://arxiv.org/abs/2505.22312).