naimulislam/reasoning_code_advanced_1m

---
license: mit
task_categories:
- text-generation
- question-answering
language:
- en
tags:
- reasoning
- chain-of-thought
- code
- synthetic
- programming
- logic
- cot
size_categories:
- 1M<n<10M
pretty_name: Reasoning Code Advanced 1M
---

# 💻 Reasoning Code Advanced 1M

## 📖 Dataset Summary
**Reasoning Code Advanced 1M** is a massive-scale, synthetic dataset specifically engineered to improve the algorithmic reasoning and problem-solving capabilities of Large Language Models (LLMs). Featuring 1,000,000 unique coding samples, this dataset spans multiple programming languages (Python, JS, C++, etc.) and focuses on logic-heavy development tasks.

A key feature of this dataset is its **Adaptive Reasoning Architecture**. Unlike standard instruction datasets, the Chain-of-Thought (CoT) traces are strictly correlated with difficulty levels and are encapsulated within `<thinking>` tags to train the model's "internal monologue" before it emits code.

## ⚙️ Dataset Structure
Each record follows a structured JSON format optimized for fine-tuning:

| Field | Type | Description |
| :--- | :--- | :--- |
| `serial_number` | int | Unique identifier for the sample. |
| `difficulty` | str | Complexity level: **Easy**, **Medium**, or **Hard**. |
| `question` | str | The coding prompt or technical challenge. |
| `reasoning` | str | The internal thought process (CoT) wrapped in `<thinking>...</thinking>`. |
| `final_answer` | str | The actual code block or solution. |

### Data Instance Example (Hard Difficulty)
```json
{
"serial_number": 88210,
"difficulty": "Hard",
"question": "Python: Write a function to check if a binary tree is height-balanced.",
"reasoning": "<thinking>To check if a tree is balanced, I need to calculate the height of left and right subtrees for every node. A tree is balanced if the absolute difference between heights is ≤ 1. I should use a recursive DFS approach. To optimize, I can return -1 if a subtree is unbalanced to avoid redundant calculations.</thinking>",
"final_answer": "def is_balanced(root):\n def check(node):\n if not node: return 0\n left = check(node.left)\n right = check(node.right)\n if left == -1 or right == -1 or abs(left - right) > 1: return -1\n return max(left, right) + 1\n return check(root) != -1"
}
```

## 🧠 Difficulty & Reasoning Logic
This dataset is designed to teach models **when** to invoke deep reasoning. The `reasoning` field population follows a strict probability distribution:

| Difficulty | Reasoning Presence | Training Objective |
| :--- | :--- | :--- |
| **Easy** | 0% (None) | Focuses on direct syntax recall and simple definitions without over-thinking. |
| **Medium** | 50% | Teaches the model to evaluate if a task requires a plan or can be solved directly. |
| **Hard** | 100% | Forces a full step-by-step logic trace before generating complex algorithms. |

## 💻 How to Use
Load the dataset via the Hugging Face `datasets` library:

```python
from datasets import load_dataset

# Replace with your actual repository path
dataset = load_dataset("naimulislam/reasoning_code_advanced_1m")

# Accessing a Hard sample
sample = dataset['train'][10]
print(f"[{sample['difficulty']}] {sample['question']}\n{sample['final_answer']}")
```

## 🛠️ Dataset Creation & Scope
The dataset was synthetically generated using a logic engine designed to simulate real-world programming scenarios across:
* **Data Structures:** Trees, Graphs, Linked Lists, and Hash Maps.
* **Algorithms:** Sorting, Searching, Dynamic Programming, and Recursion.
* **System Design:** Basic architecture patterns and class structures.
* **Debugging:** Identifying and fixing logical errors in provided snippets.
* **Optimization:** Converting O(n²) solutions into O(n log n).

## 📜 License
This dataset is released under the **MIT License**.

## 🤝 Citation
If you use this dataset in your research or LLM training, please credit:
```bibtex
@dataset{reasoning_code_advanced_1m,
author = {Naimul Islam Nahid},
title = {Reasoning Code Advanced 1M},
year = {2025},
publisher = {Hugging Face},
journal = {naimulislam/reasoning_code_advanced_1m},
}
```