OmniThought-0528

# OmniThought-0528
*A High-Quality Chain-of-Thought (CoT) Dataset for Enhanced Model Distillation*

## **Overview**
OmniThought-0528 is an advanced version of the **OmniThought** dataset, designed to enhance reasoning capabilities in language models through high-quality **Chain-of-Thought (CoT)** distillation. It consists of **365,000 reasoning chains** across diverse domains, including **mathematics, coding, and science**, generated and rigorously validated using state-of-the-art teacher models.

## **Dataset Construction**
### **Data Sources & Processing**
- **Input Problems**: Collected from multiple public sources, covering **mathematical, scientific, and programming reasoning tasks**.
- **Teacher Models**: Generated using **DeepSeek-R1** and **QwQ-32B**, then filtered via **"LLM-as-a-Judge"** validation to ensure correctness.
- **Quality Control**: Only verified reasoning chains were retained, ensuring high reliability.

### **Annotations & Metrics**
Each reasoning chain is annotated with two key metrics:

| **Score** | **Reasoning Verbosity (RV)** | **Cognitive Difficulty (CD)** |
|-----------|-----------------------------|-------------------------------|
| **0-1** | Minimal explanation, direct answers. | Elementary knowledge, single-step reasoning. |
| **2-3** | Concise reasoning with necessary steps. | Multi-step arithmetic, rule-based logic. |
| **4-5** | Detailed explanations, thorough reasoning. | Basic algebra, non-trivial inference. |
| **6-7** | Comprehensive exploration, advanced techniques. | Higher-level methods (e.g., dynamic programming, proofs). |
| **8-9** | Deep, exhaustive reasoning with nested arguments. | Highly abstract (e.g., complex algorithm analysis). |

These metrics help **match reasoning complexity with model size**—smaller models benefit from simpler chains, while larger models leverage advanced reasoning.

## **Key Features**
- **Enhanced Teacher Model**: Generated using **DeepSeek-R1-0528**, a top-tier model (comparable to **GPT-4o & Gemini-2.5-Pro**) excelling in **math, coding, and logic benchmarks**.
- **Structured Format**: Each entry includes:
```json
{
"question": "problem_text",
"reasoning": [
{
"full_response": "detailed_answer",
"Cognitive_Difficulty": score,
"Reasoning_Verbosity": score
},
...
]
}
```

# OmniThought-0528
*面向模型蒸馏优化的高质量思维链（Chain-of-Thought, CoT）数据集*

## **概述**
OmniThought-0528 是 **OmniThought** 数据集的进阶版本，旨在通过高质量思维链（Chain-of-Thought，CoT）蒸馏提升语言模型的推理能力。该数据集涵盖数学、编码与科学等多元领域的365,000条推理链，由当前顶尖的教师模型生成并经过严格验证。

## **数据集构建**
### **数据来源与处理流程**
- **输入问题**：采集自多个公开来源，覆盖数学、科学与编程推理任务。
- **教师模型**：由 DeepSeek-R1 与 QwQ-32B 生成，随后通过「LLM-as-a-Judge」验证流程进行筛选，以确保结果正确性。
- **质量管控**：仅保留经验证的推理链，保障数据集的高可靠性。

### **标注与评价指标**
每条推理链均配有两项核心指标：

| 得分区间 | 推理详细度（Reasoning Verbosity, RV） | 认知难度（Cognitive Difficulty, CD） |
|---------|--------------------------------------|--------------------------------------|
| 0-1 | 解释极简，直接给出答案 | 基础知识点，单步推理 |
| 2-3 | 包含必要步骤的简洁推理 | 多步算术、基于规则的逻辑推理 |
| 4-5 | 详细解释，推理过程完整 | 基础代数、非平凡推理 |
| 6-7 | 全面探索，运用高级技术 | 高阶方法（如动态规划、证明推导） |
| 8-9 | 深度穷尽式推理，包含嵌套论证 | 高度抽象（如复杂算法分析） |

上述指标可用于**匹配推理复杂度与模型规模**——小型模型适配更简单的推理链，而大型模型则可利用高阶推理内容。

## **核心特性**
- **增强型教师模型**：由 DeepSeek-R1-0528 生成，该模型为顶尖模型（可与 GPT-4o 及 Gemini-2.5-Pro 媲美），在数学、编码与逻辑基准测试中表现优异。
- **结构化格式**：每条数据条目包含如下内容：
json
{
"question": "problem_text",
"reasoning": [
{
"full_response": "detailed_answer",
"Cognitive_Difficulty": score,
"Reasoning_Verbosity": score
},
...
]
}

搜集汇总

背景与挑战

背景概述

OmniThought-0528是一个高质量链式思考数据集，旨在通过蒸馏提升语言模型的推理能力。它包含365,000个经过验证的推理链，覆盖数学、编码和科学等多个领域，由先进的教师模型生成并标注了认知难度和推理详细度指标。