Sidsidney/general-reasoning-ift-pairs

---
dataset_info:
features:
- name: prompt
dtype: string
- name: reasoning
dtype: string
- name: ift
dtype: string
splits:
- name: reasoning_ift_pairs
num_bytes: 12536466590
num_examples: 990098
- name: reasoning
num_bytes: 10814569775
num_examples: 990098
- name: ift
num_bytes: 2274571843
num_examples: 990098
download_size: 12132970012
dataset_size: 25625608208
configs:
- config_name: default
data_files:
- split: reasoning_ift_pairs
path: data/reasoning_ift_pairs-*
- split: reasoning
path: data/reasoning-*
- split: ift
path: data/ift-*
license: mit
task_categories:
- question-answering
- text-generation
language:
- en
tags:
- instruction-tuning
- reasoning
- synthetic
pretty_name: IFT & Reasoning Paired Dataset
size_categories:
- 1M<n<10M
---

# Reasoning-IFT Pairs (General Domain)

<p align="left">
<img src="https://cdn-avatars.huggingface.co/v1/production/uploads/62be186a5f59ff2320e6e32b/GjJ15tY7-F4bqR96FN4pd.png" alt="Dataset Icon" width="180"/>
</p>

<p align="left">
<a href="https://arxiv.org/pdf/2509.22193" target="_blank" rel="noopener noreferrer">
<img src="https://img.shields.io/badge/arXiv-2509.22193-b31b1b.svg?style=for-the-badge" alt="arXiv:2509.22193" />
</a>
</p>

This dataset provides **the largest set of IFT and Reasoning answers pairs** for a set of general domain queries (cf: [math-domain](https://huggingface.co/datasets/When-Does-Reasoning-Matter/math-reasoning-ift-pairs)).<br>
It is based on the `Infinity-Instruct` dataset, an extensive and high-quality collection of instruction fine-tuning data.

We curated **900k queries** from the `7M_core` subset of Infinity-Instruct, which covers multiple domains including general knowledge, commonsense Q&A, coding, and math.
For each query, we used [Qwen/Qwen3-235B-A22B](https://huggingface.co/Qwen/Qwen3-235B-A22B), which supports a configurable reasoning flag, to generate two answer formats:

- **IFT Answer** → concise, direct response
- **Reasoning Answer** → response with reasoning mode enabled (chain-of-thought style)

If you use this dataset in your work, please cite: **[When Does Reasoning Matter?](https://arxiv.org/pdf/2509.22193)**

```bibtex
@misc{boizard2025doesreasoningmattercontrolled,
title={When Does Reasoning Matter? A Controlled Study of Reasoning's Contribution to Model Performance},
author={Nicolas Boizard and Hippolyte Gisserot-Boukhlef and Kevin El-Haddad and Céline Hudelot and Pierre Colombo},
year={2025},
eprint={2509.22193},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2509.22193},
}
```

---

## 📂 Dataset Details

- **Source**: Based on *Infinity-Instruct* (`7M_core` subset)
- **Size**: ~900k query–answer pairs
- **Format**: Each entry contains:
- `prompt`: input question
- `reasoning`: synthetic answer with reasoning enabled
- `ift`: synthetic answer without reasoning
- **Model used for generation**: `Qwen/Qwen3-235B-A22B` (open-weight, mixture-of-experts, reasoning toggle)

---

## 🎯 Research Motivation

Frontier research initiatives highlight the potential of reasoning models, but progress is often confounded by opaque data mixtures and shifting supervision schemes.
This dataset moves the needle by isolating reasoning itself:

- Using a single teacher model to generate paired IFT and reasoning answers for the same queries, we enable clean attribution of performance improvements specifically to reasoning.
- This controlled setup avoids reliance on expensive RL pipelines (e.g. Magistral, Qwen3).
- It facilitates systematic study across model scales and data domains.

dataset_info: 数据集信息
features: 特征项
- name: 提示词（prompt）
dtype: 字符串（string）
- name: 推理过程（reasoning）
dtype: 字符串（string）
- name: 指令微调样本（ift, Instruction Fine-Tuning）
dtype: 字符串（string）
splits: 数据划分
- name: 推理-指令微调配对样本（reasoning_ift_pairs）
num_bytes: 字节数：12536466590
num_examples: 样本数量：990098
- name: 仅推理样本（reasoning）
num_bytes: 字节数：10814569775
num_examples: 样本数量：990098
- name: 仅指令微调样本（ift）
num_bytes: 字节数：2274571843
num_examples: 样本数量：990098
download_size: 下载大小：12132970012
dataset_size: 数据集总大小：25625608208
configs: 配置项
- config_name: 默认配置（default）
data_files: 数据文件
- split: 推理-指令微调配对样本（reasoning_ift_pairs）
path: data/reasoning_ift_pairs-*
- split: 仅推理样本（reasoning）
path: data/reasoning-*
- split: 仅指令微调样本（ift）
path: data/ift-*
license: 许可证：MIT许可证
task_categories: 任务类别
- 问答（question-answering）
- 文本生成（text-generation）
language: 语言：英语（en）
tags: 标签
- 指令微调（instruction-tuning）
- 推理（reasoning）
- 合成数据集（synthetic）
pretty_name: 易读名称：IFT与推理配对数据集（IFT & Reasoning Paired Dataset）
size_categories: 样本规模类别：100万<样本数<1000万

# 通用领域推理-指令微调配对数据集

<p align="left">
<img src="https://cdn-avatars.huggingface.co/v1/production/uploads/62be186a5f59ff2320e6e32b/GjJ15tY7-F4bqR96FN4pd.png" alt="数据集图标" width="180"/>
</p>

<p align="left">
<a href="https://arxiv.org/pdf/2509.22193" target="_blank" rel="noopener noreferrer">
<img src="https://img.shields.io/badge/arXiv-2509.22193-b31b1b.svg?style=for-the-badge" alt="arXiv:2509.22193" />
</a>
</p>

本数据集为通用领域查询集提供了**目前规模最大的指令微调（IFT）与推理答案配对样本集**（可对比参考数学领域数据集：[math-domain](https://huggingface.co/datasets/When-Does-Reasoning-Matter/math-reasoning-ift-pairs)）。
本数据集基于`Infinity-Instruct`数据集构建，后者是一个规模庞大、质量上乘的指令微调数据合集。

我们从`Infinity-Instruct`的`7M_core`子集中精选了**90万个查询样本**，该子集覆盖多个领域，包括常识知识、常识问答、编程与数学等。
针对每个查询，我们使用[Qwen/Qwen3-235B-A22B](https://huggingface.co/Qwen/Qwen3-235B-A22B)生成两种格式的答案，该模型支持可配置的推理开关：
- **指令微调答案**：简洁直接的应答内容
- **推理答案**：启用推理模式生成的应答（即思维链（Chain-of-Thought, CoT）风格）

若您在研究中使用本数据集，请引用：**[《推理何时发挥作用？》(When Does Reasoning Matter?)](https://arxiv.org/pdf/2509.22193)**

bibtex
@misc{boizard2025doesreasoningmattercontrolled,
title={When Does Reasoning Matter? A Controlled Study of Reasoning's Contribution to Model Performance},
author={Nicolas Boizard and Hippolyte Gisserot-Boukhlef and Kevin El-Haddad and Céline Hudelot and Pierre Colombo},
year={2025},
eprint={2509.22193},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2509.22193},
}

---

## 📂 数据集详情
- **数据来源**：基于*Infinity-Instruct*的`7M_core`子集
- **数据集规模**：约90万个查询-应答配对样本
- **数据格式**：每条样本包含以下字段：
- `提示词（prompt）`：输入问题
- `推理过程（reasoning）`：启用推理模式生成的合成应答
- `指令微调样本（ift）`：未启用推理模式生成的合成应答
- **生成所用模型**：`Qwen/Qwen3-235B-A22B`（开源权重、混合专家架构、支持推理开关）

---

## 🎯 研究动机
前沿研究成果凸显了推理模型的应用潜力，但现有进展常因不透明的数据集混合方式与多变的监督机制而难以归因。
本数据集通过将推理本身作为独立变量，突破了这一研究瓶颈：
1. 使用单一教师模型为同一查询生成配对的指令微调答案与推理答案，可精准将模型性能提升归因于推理模块的作用。
2. 这种可控的实验设置无需依赖成本高昂的强化学习流水线（如Magistral、Qwen3系列模型）。
3. 可支持针对不同模型规模与数据领域的系统性研究。