Corpus-NZ/CoT-for-LLM

README — Advanced Chain‑of‑Thought Dataset Generator

Overview

This project generates a large-scale synthetic dataset of Chain‑of‑Thought (CoT) reasoning examples across multiple domains:

• —Math (algebra, word problems, multi‑step reasoning)
• —English (vocabulary explanations, nuance, tone)
• —Writing (multi‑paragraph reflections, structured planning)
• —Coding (advanced algorithms, data structures, real code snippets)
• —Science (physics, biology, chemistry, earth science, etc.)

The generator produces 100,000 rows of richly varied reasoning examples designed to help train or fine‑tune language models to produce human‑like, multi‑step reasoning.

Each row includes:

• —id
• —category
• —difficulty
• —prompt
• —reasoning
• —final_answer

Key Features

1. Difficulty Levels

Every example is tagged with one of:

• —easy
• —medium
• —hard
• —expert

Difficulty affects:

• —math complexity
• —reasoning depth
• —coding task difficulty
• —writing topic depth
• —science concept complexity

2. Human‑Like Reasoning Styles

The generator uses four distinct reasoning styles:

• —Linear — clean, numbered steps
• —Messy — human‑like hesitations and corrections
• —Two‑Methods — compares two different solution paths
• —Long‑Paragraph — narrative, reflective reasoning

This prevents robotic patterns and increases realism.

3. Large Topic Pools

Each domain pulls from expanded lists:

• —15+ math word‑problem types
• —25+ English vocabulary words
• —15+ writing themes
• —15+ advanced coding tasks
• —20+ science concepts

This ensures variety across 100k rows.

4. Advanced Coding Tasks + Real Code Snippets

Coding examples include:

• —Dijkstra’s algorithm
• —Merge sort
• —JSON parsing
• —Binary search trees
• —Graph cycle detection
• —LRU cache
• —Expression parsing

Each example includes a real code snippet in Python, JavaScript, or pseudocode.

5. Harder Math + Realistic Word Problems

Math examples include:

• —multi‑step algebra
• —two‑variable equations
• —distance/time/speed
• —probability
• —geometry
• —compound interest
• —mixture problems

Word problems use real‑world contexts.

6. Multi‑Paragraph Reasoning Chains

Writing and science examples often produce long‑form reasoning, not just bullet points.