OpenO1-SFT-Ultra

# openo1-sft-ultra-35m-data

## Instruction
We have released the openo1-sft-ultra-35m-data, which contains 35 million data points. It is based on existing open-source datasets and synthesized using the openo1-qwen-sft model. We first collected open-source datasets and then annotated the data based on difficulty, quality, and question types using the qwen-2.5-72b-instruct model. To ensure the difficulty and quality of the data, we only retained data where both the difficulty and quality are ≥8.

## Data format
- 'uid': Data ID
- 'query': Original data query
- 'response': Long COT response, including detailed thought process
- 'source': Data source
- 'difficulty': Question difficulty, range from 1 to 10
- 'quality': Data quality, range from 1 to 10
- 'answer': Ground truth answer of the data
- 'query_len': Length of the question
- 'response_len': Length of the answer
- 'Topic': Data topic category, including Math | Code | Reasoning
- 'answer_type': Answer type annotation
- For math domain:
- a: Purely numerical answer
- b: Purely formulaic answer
- c: Long textual explanation
- For code domain:
- a: Answer that includes code
- b: Answer that contains only code-related text
- For reasoning domain:
- a: Answer consisting of only words/phrases
- b: Answer that includes a long textual explanation
- "cases": Test cases included in the code

## Data statistics
### Data source
We have used the following sources for the data:
- WebInstructFull
- homework
- infinity-instruct
- math-stack-exchange
- MathInstruct
- mcq


<!-- Insert statistical charts for different sources here -->

# openo1-sft-ultra-35m-data

## 数据集说明
我们现已发布openo1-sft-ultra-35m-data数据集，该数据集共包含3500万条数据样本。本数据集基于现有开源数据集构建，并通过openo1-qwen-sft模型生成合成数据。我们首先收集开源数据集，随后使用qwen-2.5-72b-instruct模型，基于难度、质量与题型对数据进行标注。为保障数据的难度与质量，我们仅保留难度与质量评分均≥8的样本。

## 数据格式
- `'uid'`：数据ID
- `'query'`：原始数据查询内容
- `'response'`：长链式思考（Chain of Thought, COT）响应，包含完整推理过程
- `'source'`：数据来源
- `'difficulty'`：题目难度评分，取值范围为1至10
- `'quality'`：数据质量评分，取值范围为1至10
- `'answer'`：数据集的标准答案（Ground Truth Answer）
- `'query_len'`：问题文本长度
- `'response_len'`：响应文本长度
- `'Topic'`：数据主题类别，涵盖数学（Math）、代码（Code）与推理（Reasoning）三大领域
- `'answer_type'`：答案类型标注，按领域细分如下：
- 数学领域：
- a：纯数值型答案
- b：纯公式型答案
- c：长文本解释型答案
- 代码领域：
- a：包含代码的答案
- b：仅包含代码相关文本的答案
- 推理领域：
- a：仅由单词/短语组成的答案
- b：包含长文本解释的答案
- `"cases"`：代码中包含的测试用例

## 数据统计
### 数据来源
本数据集采用以下数据源：
- WebInstructFull
- homework
- infinity-instruct
- math-stack-exchange
- MathInstruct
- mcq

<!-- 此处插入不同数据源的统计图表 -->