README.md

---
language:
- en
license: apache-2.0
size_categories:
- n<1K
task_categories:
- text-generation
dataset_info:
  features:
  - name: source
    dtype: string
  - name: question
    dtype: string
  - name: answer
    dtype: string
  - name: constraint_desc
    list: string
  - name: key
    dtype: string
  - name: instruction_id_list
    list: string
  - name: kwargs
    list:
    - name: capital_frequency
      dtype: int64
    - name: capital_relation
      dtype: string
    - name: num_words
      dtype: int64
    - name: relation
      dtype: string
    - name: keyword
      dtype: string
    - name: frequency
      dtype: int64
    - name: prompt_to_repeat
      dtype: string
    - name: keywords
      list: string
    - name: forbidden_words
      list: string
    - name: num_highlights
      dtype: int64
    - name: end_phrase
      dtype: string
    - name: num_bullets
      dtype: int64
    - name: section_spliter
      dtype: string
    - name: num_sections
      dtype: int64
    - name: language
      dtype: string
  - name: prompt
    dtype: string
  splits:
  - name: dev
    num_bytes: 123575
    num_examples: 90
  - name: test
    num_bytes: 478304
    num_examples: 332
  download_size: 223359
  dataset_size: 601879
configs:
- config_name: default
  data_files:
  - split: dev
    path: data/dev-*
  - split: test
    path: data/test-*
---

# Math-IF Dataset Card

This dataset is associated with the paper [From Leaky Thoughts to Private Reasoning: Controlling What LRMs Say to Themselves](https://huggingface.co/papers/2602.24210).

The official code repository for the project is available here: [UKPLab/arxiv2026-controllable-reasoning-models](https://github.com/UKPLab/arxiv2026-controllable-reasoning-models).

## Dataset Description

Math-IF (MathIF) is an instruction-following benchmark built on top of math word problems. Each example includes a math question together with explicit, verifiable instructions about how the model should respond (e.g., format, style, or structural constraints). The benchmark is designed to jointly test:

- instruction following in the **reasoning trace (RT)** and  
- instruction following and correctness in the **final answer (FA)**.

In this repository, Math-IF is used as both a development set and a test benchmark for controllable reasoning models.

## Intended Use

- Evaluate how well models follow explicit instructions when solving math problems.

The dataset is intended for **research and benchmarking** only.

## Dataset Structure

From the accompanying paper:

- **Size**:
  - **Dev**: 90 examples  
  - **Test**: 332 examples  
- **Splits used here**:
  - The **GSM8K partition** is used as dev set for model selection.
  - The remaining partition is used as test set.

Each instance conceptually includes:

- **`prompt`**: the user prompt with the math question and instruction.
- **`answer`**: the ground-truth final answer.
- **`question`**: the underlying math word problem (without instructions).
- **metadata for evaluation**: information needed to compute instruction-following metrics and answer accuracy.

## Tasks and Evaluation

- **Main task**: Instruction-following on math problems.
- **Metrics**:
  - *Instruction-level loose-accuracy* (as defined in the Math-IF paper) for both RTs and FAs, yielding **IF-RT** and **IF-FA**.
  - **Answer accuracy** measuring whether the final numeric answer is correct.

## Data Source

Math-IF was introduced to study the trade-off between reasoning performance and instruction-following in large reasoning models. For complete details, examples, and official evaluation scripts, please see the original Math-IF paper and repository.

## License

- **License**: Apache 2.0  

## Known Limitations and Considerations

- The dataset focuses on **math word problems**, so instruction-following performance may differ on other domains (e.g., open-ended dialogue, code generation).
- The benchmark size is modest (422 examples total in the dev+test configuration used here), which can make very fine-grained comparisons noisy.
- Instructions are in **English**, so the benchmark does not directly evaluate multilingual behavior.

## Citation

```bibtex
@misc{puerto2026controllablereasoningmodelsprivate,
      title={Controllable Reasoning Models Are Private Thinkers}, 
      author={Haritz Puerto and Haonan Li and Xudong Han and Timothy Baldwin and Iryna Gurevych},
      year={2026},
      eprint={2602.24210},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2602.24210}, 
}

@article{fu2025scaling,
  title={Scaling Reasoning, Losing Control: Evaluating Instruction Following in Large Reasoning Models},
  author={Fu, Tingchen and Gu, Jiawei and Li, Yafu and Qu, Xiaoye and Cheng, Yu},
  journal={arXiv preprint arXiv:2505.14810},
  year={2025}
}
```