Demeter-LongCoT-400K


# **Demeter-LongCoT-400K**

> **Demeter-LongCoT-400K** is a high-quality, compact chain-of-thought reasoning dataset curated for tasks in mathematics, science, and coding. While the dataset spans diverse domains, it is primarily driven by mathematical reasoning, reflecting a major share of math-focused prompts and long-form logical solutions.

## Quick Start with Hugging Face Datasets🤗

```py
pip install -U datasets
```

```py
from datasets import load_dataset

dataset = load_dataset("prithivMLmods/Demeter-LongCoT-400K", split="train")
```

## Overview

* **Total Samples**: \~400,000
* **Split**: `train` only
* **Languages**: English
* **Format**: Apache Arrow (auto-converted to Parquet)
* **License**: Apache-2.0
* **Tags**: `math`, `code`, `science`, `reasoning`, `longcot`

## Highlights

* Structured to promote **long-form, step-by-step reasoning**, ideal for training and evaluating chain-of-thought (CoT) capable models.
* Reasoning traces include natural, human-like explanations for both simple and complex problems.
* Fine-tuned across math word problems, logic-based questions, and technical prompts from STEM domains.

## Dataset Structure

Each entry in the dataset includes:

* **`problem`** (string): A math, science, or code problem.
* **`solution`** (string): A detailed step-by-step solution crafted in a long-form reasoning style.

The reasoning structure in solutions helps models understand logical flow, intermediate steps, and layered deductions—making this dataset suitable for advanced LLMs requiring interpretable outputs.

## Source & Derivation

Demeter-LongCoT-400K is a **random seed subset** derived from:

* [Demeter-LongCoT-6M](https://huggingface.co/datasets/prithivMLmods/Demeter-LongCoT-6M) (\~6.4M samples).
* Curated to \~400K entries while maintaining diverse coverage across domains.
* Generated from a custom internal modular dataset tailored for logical and numeric reasoning tasks, with chain-of-thought style responses produced by QwQ 32B-based models.

This dataset was created with a focus on enhancing CoT capabilities in large-scale models working on math, science, and code.

## License

Apache License 2.0


# **Demeter-LongCoT-400K**

> **Demeter-LongCoT-400K** 是一款高质量、轻量化的思维链（chain-of-thought, CoT）推理数据集，专为数学、科学与编码类任务打造。尽管该数据集覆盖多元领域，但核心以数学推理为主，占比最高的提示词与长格式逻辑解答均围绕数学场景展开。

## Hugging Face 数据集快速入门🤗

py
pip install -U datasets

py
from datasets import load_dataset

dataset = load_dataset("prithivMLmods/Demeter-LongCoT-400K", split="train")

## 数据集概览

* **总样本量**: 约400,000
* **数据集拆分**: 仅包含训练集（`train`）
* **语言**: 英语
* **数据格式**: Apache Arrow格式（自动转换为Parquet格式）
* **许可证**: Apache-2.0
* **标签**: `数学`, `代码`, `科学`, `推理`, `longcot`

## 数据集亮点

* 数据集设计旨在促进**长格式、逐步式推理**，非常适合训练与评估具备思维链（chain-of-thought, CoT）能力的模型。
* 推理轨迹涵盖针对简单与复杂问题的自然类人解释。
* 数据集覆盖STEM领域的数学应用题、逻辑推理题与技术类提示词，并完成了微调适配。

## 数据集结构

数据集中的每条样本包含以下字段：
* **`problem`**（字符串类型）：数学、科学或编码类问题。
* **`solution`**（字符串类型）：采用长格式推理风格编写的详细分步解答。

解答中的推理结构可帮助模型理解逻辑流程、中间步骤与层级演绎，因此该数据集非常适合需要可解释输出的高级大语言模型（Large Language Model, LLM）。

## 数据集来源与衍生说明

Demeter-LongCoT-400K 是从以下数据集抽取的**随机种子子集**：
* [Demeter-LongCoT-6M](https://huggingface.co/datasets/prithivMLmods/Demeter-LongCoT-6M)（约640万条样本）。
* 经筛选后保留约40万条样本，同时维持各领域的覆盖多样性。
* 数据集源自专为逻辑与数值推理任务定制的内部模块化数据集，其思维链风格的回复由基于QwQ 32B的模型生成。

本数据集的构建目标是提升面向数学、科学与编码任务的大规模模型的思维链能力。

## 许可证

Apache许可证2.0