wflying/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, 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

Distribution by number of constraints:

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:

{
  "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. 1.Validate the structure and required non-empty values of the JSON record, prompt, metadata, instruction IDs, and kwargs.
2. 2.Verify that every instruction ID maps to a defined constraint checker.
3. 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. 4.Reconstruct the natural-language constraint description and verify that it agrees with both the prompt text and kwargs.
5. 5.Remove known mismatches between prompt descriptions and checker behavior, invalid copy-span boundaries, and deterministic cross-constraint conflicts.
6. 6.Deduplicate records by the SHA-256 hash of the normalized prompt and write the output atomically through a temporary file.

Loading the dataset

from datasets import load_dataset

dataset = load_dataset(
    "wflying/instruction-following-rl-66k",
    split="train",
    streaming=True,
)

sample = next(iter(dataset))
print(sample["prompt"])
print(sample["metadata"]["instruction_id_list"])

License and responsible use

The upstream dataset is licensed under ODC-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. 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

Please cite the upstream work:

@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}
}