Datasets:

Nebulixlabs
/

Legend-Math

license: apache-2.0
language:
  - en
task_categories:
  - text-generation
  - question-answering
tags:
  - math
  - synthetic
  - instruction-tuning
  - sft
  - arithmetic
  - catastrophic-forgetting-prevention
size_categories:
  - 10M<n<100M

🏆 Legend-Math Dataset

Welcome to the Legend-Math dataset! This is a massive, highly optimized, and meticulously generated dataset containing 10 Million rows of diverse mathematical problems. It is designed to transform any Large Language Model (LLM) into a mathematical genius through Supervised Fine-Tuning (SFT).

📊 Dataset Overview

Name: Legend-Math
Size: 10,000,000 (10 Million) Records
Total Tokens: ~350 Million Tokens (Approx. 35 tokens per row)
Format: JSONL (OpenAI ChatML / Messages format)
Language: English / Universal Math
Use Case: Supervised Fine-Tuning (SFT), Continuous Pre-Training (CPT)

🧠 What Will the Model Learn? (Capabilities)

By fine-tuning your model on Legend-Math, it will master:

Basic Arithmetic: High-accuracy Addition, Subtraction, Multiplication, Division, and Modulo operations (+, -, *, /, %).
Multi-Step Equations: Solving complex BODMAS/PEMDAS expressions systematically.
Advanced Mathematics:
- Exponential calculations and powers (e.g., $x^y$).
- Exact Geometry calculations using Pi (π).
- High-precision Square Roots.
- Trigonometry (Sine, Cosine, Tangent).
Precision & Accuracy: The answers are generated via Python's execution engine, ensuring 0% hallucination in the dataset's ground truth.

🚀 Model SFT Capability (How large a model can you train?)

With ~350 Million high-quality instruction-response tokens, this dataset is powerful enough to fine-tune models of virtually any size:

Small Models (1B - 3B): Perfect for making edge-device math assistants (e.g., TinyLlama, Qwen-1.5-1.8B).
Medium Models (7B - 14B): Ideal for robust SFT on models like Llama-3-8B, Mistral-7B, or Qwen-7B. It will easily push their math benchmarks (GSM8K, MATH) to top-tier levels.
Large Models (30B - 70B+): You can safely use this for continuous pre-training or aggressive SFT on models like Llama-3-70B or Mixtral to create a specialized, world-class Math Oracle.

🛡️ Why Choose Legend-Math? (The Secret Sauce)

1. Zero Catastrophic Forgetting 🧠💡

One of the biggest problems with fine-tuning a model heavily on Math is Catastrophic Forgetting—the model becomes great at math but forgets how to hold a normal conversation or loses its natural language charm. Legend-Math solves this natively! Every 50,000th row in this dataset injects a high-quality programming joke. This periodic language-shift acts as a memory refresher, ensuring the model retains its conversational skills, humor, and general language abilities while mastering math.

2. Ready-to-Train Format 🛠️

No data wrangling required. The dataset is already formatted in the exact schema expected by modern fine-tuning libraries (Axolotl, LLaMA-Factory, Unsloth, Hugging Face TRL).

3. Computationally Perfect 🎯

Unlike scraped datasets that contain human errors or LLM hallucinations, every single math problem here was dynamically generated and computed.

📂 Dataset Structure

Each record in the dataset is structured as follows:

{
  "id": "6790b369d262461abba8c3f3de1425d9",
  "date": "2026-06-30 09:29:40",
  "tokens": 14,
  "messages": [
    {
      "role": "user",
      "content": "Calculate the exact area of a circle with radius 45. Use pi."
    },
    {
      "role": "assistant",
      "content": "6361.7251"
    }
  ]
}