rl-if-math-coding-logic-60k

# RL IF Math Coding Logic 60K This dataset is a mixed training set for reinforcement learning with verifiable rewards. It contains 60,680 JSON Lines sampled and combined from four public datasets covering instruction following, mathematics, code, and logic/table reasoning. There is one `train` split and no validation or test split. The release is intended for research and development. It is not a benchmark, and downstream users should construct their own held-out evaluations and check for contamination before using it to report comparable results. ## Source composition | Category | `metadata.data_type` | Examples | Source dataset | Source license | |---|---:|---:|---|---| | Instruction following | `if` | 15,679 | [allenai/IF_multi_constraints_upto5](https://huggingface.co/datasets/allenai/IF_multi_constraints_upto5) | ODC-BY-1.0 | | Mathematics | `klear_math` | 15,000 | [Kwai-Klear/KlearReasoner-MathSub-30K](https://huggingface.co/datasets/Kwai-Klear/KlearReasoner-MathSub-30K) | Apache-2.0 | | Code | `klear_code` | 15,001 | [Kwai-Klear/KlearReasoner-CodeSub-15K](https://huggingface.co/datasets/Kwai-Klear/KlearReasoner-CodeSub-15K) | Apache-2.0 | | Logic and table reasoning | `nemotron_rl_reasoning_gym` | 7,500 | [wflying/logic-rl-24k](https://huggingface.co/datasets/wflying/logic-rl-24k) | CC BY 4.0 | | Logic and table reasoning | `guru_rl` | 7,500 | [wflying/logic-rl-24k](https://huggingface.co/datasets/wflying/logic-rl-24k) | MIT | The logic/table categories are routed through `wflying/logic-rl-24k`; that repository documents its own upstream attribution to NVIDIA Nemotron RL Reasoning Gym and LLM360 Guru RL. ## File and integrity information - File: `rl-if-math-coding-logic-60k.jsonl` - Format: UTF-8 JSON Lines, one JSON object per line - Lines: 60,680 - Size: 8,090,516,042 bytes - SHA-256: `d37fbaa14a2c7da1072805ffd4c22c438d8ac12c5eb4f16e3ff4d1354e8a13dc` ## Data structure Each record has three top-level fields: ```json { "prompt": [ {"role": "user", "content": "..."} ], "label": "...", "metadata": {"data_type": "..."} } ``` - `prompt`: a chat-formatted user prompt, normally containing one user message. - `label`: the reference answer, executable test cases, or an empty string for instruction-following examples whose reward is computed from the prompt constraints. - `metadata`: source, reward-routing, ground-truth, constraint, token-count, and task-specific auxiliary fields. Its nested schema is intentionally heterogeneous. Important common metadata fields include: - `data_type`: top-level task/reward family used to route the example. - `ground_truth`: expected answer where applicable. - `num_tokens`: precomputed token count where supplied by the source pipeline. - `instruction_id_list` and `instruction_kwargs`: verifiable instruction constraints for `if` records. - `data_source`, `ability`, and `style`: source family and reward style for several math, code, and logic records. Because `label` and parts of `metadata` have mixed JSON types, preserve them as JSON when loading into a fixed-schema framework. For example, JSON-encode those fields before constructing a columnar table. ## Loading Load the dataset with ModelScope: ```python from modelscope.msdatasets import MsDataset dataset = MsDataset.load( "wflying000/rl-if-math-coding-logic-60k", split="train", ) ``` Users who require `datasets.Dataset` should first normalize heterogeneous JSON fields or define an explicit schema appropriate for their training stack. ## Intended use The dataset is intended for: - reinforcement learning with rule-based or verifiable rewards; - multi-task training across instruction following, mathematics, code, logic, and table reasoning; - reward-function routing experiments based on `metadata.data_type`; - controlled studies of mixed-domain training mixtures. Reward implementations must respect the expected output format in each prompt. These formats include, among others, boxed mathematics answers, code test cases, `<answer>...</answer>` tags, and instruction-following constraints. ## Limitations - This is only a training mixture; no held-out split is provided. - Task families and answer formats are not uniformly distributed. - `label` and `metadata` are heterogeneous and may require normalization. - Reference answers and generated source data can contain errors, ambiguities, bias, or duplicate concepts. - The compilation has not been independently audited for personal information, harmful content, or downstream benchmark contamination. - Instruction-following examples may include content generated by third-party models subject to separate terms. Users are responsible for validating task suitability, licenses, reward functions, and train/evaluation separation for their intended application. ## Licensing and attribution The repository uses the Hub's `other` license tag because it combines records under multiple licenses: - IF records: [ODC-BY-1.0](https://opendatacommons.org/licenses/by/1-0/), with Ai2's Responsible Use Guidelines and third-party model terms where applicable. - Klear math and code records: Apache-2.0. - `nemotron_rl_reasoning_gym` logic records: CC BY 4.0. - `guru_rl` logic/table records: MIT. Consult the linked source datasets and their upstream licenses for the exact terms and attribution requirements that apply to your use case. This card is informational and is not legal advice. ## Source citations Please cite the source datasets and papers used to create this mixture. Key source references include: ```bibtex @misc{pyatkin2025generalizing, title={Generalizing Verifiable Instruction Following}, author={Valentina Pyatkin and Saumya Malik and Victoria Graf and Hamish Ivison and Shengyi Huang and Pradeep Dasigi and Nathan Lambert and Hannaneh Hajishirzi}, year={2025}, archivePrefix={arXiv} } ``` ```bibtex @article{su2025klearreasoner, title={Klear-Reasoner: Advancing Reasoning Capability via Gradient-Preserving Clipping Policy Optimization}, author={Zhenpeng Su and Leiyu Pan and Xue Bai and Dening Liu and Guanting Dong and Jiaming Huang and Wenping Hu and Fuzheng Zhang and Kun Gai and Guorui Zhou}, journal={arXiv preprint arXiv:2508.07629}, year={2025} } ``` For logic/table records, also cite the attribution and upstream papers listed in [wflying/logic-rl-24k](https://huggingface.co/datasets/wflying/logic-rl-24k), including the Guru RL and NVIDIA Nemotron RL Reasoning Gym releases.