instruction-following-rl-66k

# Instruction Following RL 66K ## Dataset overview `instruction-following-rl-66k` is an English training dataset for instruction-following reinforcement learning (RL/RLVR), containing **66,418** examples. It is derived primarily from AllenAI's [IF_multi_constraints_upto5](https://huggingface.co/datasets/allenai/IF_multi_constraints_upto5), whose instructions contain up to five verifiable constraints drawn from IFEval and IFBench-Train. Each record is first validated for its JSON, prompt, and metadata structure. A predefined instruction-ID-to-constraint-checker mapping is then used to verify that every instruction ID has a corresponding checker and that the kwargs, checker-generated natural-language constraint description, and prompt are mutually consistent. The processing subsequently removes known mismatches between prompt descriptions and checker behavior, invalid copy-span boundaries, deterministic cross-constraint conflicts, and unstable targets that would require randomized default values to repair. Finally, records are deduplicated using the SHA-256 hash of the normalized prompt. ## Dataset statistics | Item | Value | | :--- | ---: | | Split | `train` | | Examples | 66,418 | | JSONL file size | 85,620,970 bytes (approximately 81.65 MiB) | | Data type | `if` | | Unique instruction IDs | 51 | | Constraints per example | 1–5 | | Prompt length in characters | min 35; median 638; P95 1,975; max 10,174 | | `num_tokens` | min 19; median 158; P95 523; max 2,048 | Distribution by number of constraints: | Constraints per example | Examples | Share | | ---: | ---: | ---: | | 1 | 13,428 | 20.22% | | 2 | 17,914 | 26.97% | | 3 | 17,798 | 26.80% | | 4 | 12,736 | 19.18% | | 5 | 4,542 | 6.84% | ## Data format Each line is a JSON object with the following fields: - `prompt`: A list containing a single user message in the form `[{"role": "user", "content": ...}]`. - `label`: An empty-string placeholder. Compliance is determined from the verifiable constraints stored in `metadata`; no supervised reference answer is provided. - `metadata.instruction_id_list`: The list of instruction/checker IDs to validate. - `metadata.instruction_kwargs`: A list of parameter dictionaries aligned one-to-one with `instruction_id_list`; checkers without parameters may use `null`. - `metadata.num_tokens`: A precomputed token count. The tokenizer used to produce this value is not recorded in the released file. - `metadata.data_type`: `if` for every example. Illustrative example: ```json { "prompt": [ { "role": "user", "content": "<task followed by one to five verifiable constraints>" } ], "label": "", "metadata": { "instruction_id_list": ["<checker-id-1>", "<checker-id-2>"], "instruction_kwargs": [{"<arg>": "<value>"}, null], "num_tokens": 128, "data_type": "if" } } ``` ## Data processing The dataset undergoes strict semantic validation and prompt-level deduplication: 1. Validate the structure and required non-empty values of the JSON record, prompt, metadata, instruction IDs, and kwargs. 2. Verify that every instruction ID maps to a defined constraint checker. 3. Confirm that each checker accepts the supplied kwargs keys and values as-is, rejecting invalid parameters that depend on randomized defaults for repair. 4. Reconstruct the natural-language constraint description and verify that it agrees with both the prompt text and kwargs. 5. Remove known mismatches between prompt descriptions and checker behavior, invalid copy-span boundaries, and deterministic cross-constraint conflicts. 6. Deduplicate records by the SHA-256 hash of the normalized prompt and write the output atomically through a temporary file. ## Loading the dataset ```python from modelscope.msdatasets import MsDataset dataset = MsDataset.load( "wflying000/instruction-following-rl-66k", subset_name="default", split="train", ) sample = dataset[0] print(sample["prompt"]) print(sample["metadata"]["instruction_id_list"]) ``` ## License and responsible use The upstream dataset is licensed under [ODC-BY-1.0](https://opendatacommons.org/licenses/by/1-0/). This dataset retains the same `odc-by` license metadata. Users should attribute the upstream dataset and its authors and comply with the Allen Institute for AI [Responsible Use Guidelines](https://allenai.org/responsible-use). The upstream Dataset Card additionally notes that the dataset may contain generated data subject to separate terms of use from third-party model providers; users are responsible for reviewing and complying with all applicable terms. ## Sources and citation - Upstream dataset: [allenai/IF_multi_constraints_upto5](https://huggingface.co/datasets/allenai/IF_multi_constraints_upto5) Please cite the upstream work: ```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}, eprint={TODO}, archivePrefix={arXiv}, primaryClass={cs.CL} } ```