math-rlvr-15k

# math-rlvr-15k Math RLVR 15K is a 15,000-example English mathematical reasoning dataset prepared for reinforcement learning with verifiable rewards (RLVR). It is derived from mathematical data in [allenai/RLVR-MATH](https://huggingface.co/datasets/allenai/RLVR-MATH) and [allenai/Dolci-RL-Zero-Mix-7B](https://huggingface.co/datasets/allenai/Dolci-RL-Zero-Mix-7B). Each example contains a single user prompt, a reference final answer, and compact metadata for reward routing. Prompts ask the model to place its final answer in `\boxed{}`. The dataset contains problems and answer targets; it does not contain model-generated responses or chain-of-thought traces. ## Dataset summary | Property | Value | |---|---:| | Examples | 15,000 | | Split | `train` only | | File | `math-rlvr-15k.jsonl` | | Format | UTF-8 JSON Lines; one object per line | | File size | 7,569,123 bytes (approximately 7.22 MiB) | | Unique prompts | 15,000 | | Prompt shape | One `user` message per example | | Prompt length in characters | min 73; median 308; P95 669; max 4,087 | | `metadata.num_tokens` | min 31; median 107; P95 219; max 1,773 | All 15,000 lines parse as JSON and have the same top-level and metadata schemas. Every reference answer is a non-empty string, and every normalized prompt is unique within the released file. ## Upstream sources - [allenai/RLVR-MATH](https://huggingface.co/datasets/allenai/RLVR-MATH): mathematical prompts with verifiable ground-truth answers. - [allenai/Dolci-RL-Zero-Mix-7B](https://huggingface.co/datasets/allenai/Dolci-RL-Zero-Mix-7B): a broader RLVR mixture containing math, code, instruction-following, and general data; this release uses mathematical data from that mixture. The released schema does not retain a per-example source field, so the contribution of each upstream repository cannot be reconstructed from `math-rlvr-15k.jsonl` alone. `metadata.data_type` is a reward-routing identifier, not an upstream-source identifier. ## Data format Each line follows this structure: ```json { "prompt": [ { "role": "user", "content": "<mathematics problem>\n\nPlease put your final answer in \\boxed{}." } ], "label": "<reference final answer>", "metadata": { "ground_truth": "<reference final answer>", "data_type": "klear_math", "num_tokens": 128 } } ``` ### Fields - `prompt`: A list containing exactly one user message. All prompts request a final answer in LaTeX `\boxed{}` form. - `label`: The reference final answer as a string. - `metadata.ground_truth`: The verifier target. It is identical to `label` for all 15,000 examples. - `metadata.data_type`: The reward-routing identifier; it is `klear_math` for every example. - `metadata.num_tokens`: A precomputed prompt token count. The tokenizer used to compute this value is not recorded in the released file. ## Loading the dataset ```python from modelscope.msdatasets import MsDataset dataset = MsDataset.load( "wflying000/math-rlvr-15k", split="train", ) example = dataset[0] prompt = example["prompt"] expected_answer = example["metadata"]["ground_truth"] ``` For RLVR training, pass the content of `prompt` to the model and compare the extracted boxed final answer against `metadata.ground_truth` with an appropriate mathematical answer verifier. ## Data quality and validation - 15,000 of 15,000 lines are valid JSON. - Every record contains `prompt`, `label`, and `metadata`. - Every prompt contains exactly one message with `role: "user"`. - Every prompt includes the requested `\boxed{}` answer format. - Every `label` and `metadata.ground_truth` value is a non-empty string. - `label == metadata.ground_truth` for every record. - Every `metadata.data_type` value is `klear_math`. - All 15,000 normalized prompts are unique. These checks establish structural consistency, not the mathematical correctness of every problem or reference answer. ## Intended use This dataset is intended for research and development involving: - reinforcement learning with verifiable mathematical rewards; - mathematical reasoning post-training; - evaluation of answer extraction and mathematical verification; - controlled experiments with a compact, uniform RLVR schema. ## Limitations - The dataset contains only a training split. - Per-example upstream provenance is not retained in the released schema. - `metadata.num_tokens` cannot be reproduced exactly without the original tokenizer configuration. - Exact-string comparison may reject mathematically equivalent answers; use a robust mathematical verifier when appropriate. - Source data may contain incorrect answers, ambiguous questions, malformed notation, duplicates not captured by exact prompt matching, or overlap with downstream evaluations. - This compilation has not been independently audited for personal information, harmful content, or benchmark contamination. Users should validate answer coverage, reward behavior, and train/evaluation separation for their application. ## License and attribution The [Dolci RL-Zero Mix](https://huggingface.co/datasets/allenai/Dolci-RL-Zero-Mix-7B) Dataset Card declares the Open Data Commons Attribution License v1.0 (ODC-By) and points users to the Allen Institute for AI [Responsible Use Guidelines](https://allenai.org/responsible-use). At the time this card was prepared, the [RLVR-MATH](https://huggingface.co/datasets/allenai/RLVR-MATH) repository did not declare a license in its Dataset Card or Hub metadata. Because this release combines material from both repositories, no single license identifier is asserted in this card's YAML metadata. Users are responsible for reviewing and complying with the terms, attribution requirements, and responsible-use guidance applicable to each upstream source. This section is informational and is not legal advice. ## Citation Please cite and attribute the upstream dataset repositories used in this release: - [allenai/RLVR-MATH](https://huggingface.co/datasets/allenai/RLVR-MATH) - [allenai/Dolci-RL-Zero-Mix-7B](https://huggingface.co/datasets/allenai/Dolci-RL-Zero-Mix-7B) The Dolci RL-Zero Mix Dataset Card states that its technical manuscript is forthcoming.