razielAI/HE-Math

---
license: apache-2.0
task_categories:
- text-generation
language:
- he
tags:
- math
---
# Orion-1 Math Dataset: K-12 Hebrew Educational Corpus

## Overview
The **Orion-1 Math Dataset** is a comprehensive, high-quality synthetic and curated dataset specifically engineered for training Large Language Models (LLMs) in the field of mathematics education. It is the core data foundation for **Orion-1**, a Socratic Hebrew-native model developed by **TopAI Projects**.

Unlike generic math datasets, Orion-1 is built to reflect the Israeli Ministry of Education's curriculum, focusing on pedagogical progression from elementary school to advanced high school levels.

## Dataset Structure
The dataset is meticulously organized into **12 distinct files**, each representing a full academic year in the Israeli K-12 system:

- **Elementary School (Grades 1-6):** Focuses on foundational arithmetic, number sense, the decimal system, basic geometry, fractions, and decimals.
- **Middle School (Grades 7-9):** Introduces algebraic thinking, functions, linear equations, basic statistics, and spatial geometry.
- **High School (Grades 10-12):** Advanced mathematics including Calculus (differentiation and integration), Trigonometry, Vectors, Complex Numbers, and Probability, tailored for the "Bagrut" (Matriculation) exams.

## Data Format & Schema
To ensure optimal SFT (Supervised Fine-Tuning) performance, every data entry follows a strict Socratic structure. This forces the model to learn reasoning and guidance rather than just providing immediate answers.

### The Orion Block Format:
- **Topic:** The specific mathematical concept.
- **Definition:** A simplified, age-appropriate explanation.
- **Logical Explanation:** Intuitive analogies (e.g., using money, shapes, or real-world objects).
- **Solved Example:** A step-by-step solution rendered in full LaTeX.
- **Common Error Analysis:** Identification of frequent student misconceptions and why they occur.
- **Orion Tutor Dialogue:** A Socratic interaction where the AI guides a "stuck" student through hints and leading questions without revealing the solution.

## Technical Specifications
- **Language:** Native Hebrew (Israeli dialect).
- **Mathematical Notation:** Standard LaTeX throughout all files for seamless rendering in modern UI/UX environments.
- **Total Files:** 12 (.txt or .jsonl format).
- **Pedagogical Goal:** To reduce "answer-grabbing" behavior and promote "thinking-native" AI interactions in Hebrew.

## Mission Statement
The goal of this dataset is to empower the **Duchifat model series** to become the most advanced educational AI in Israel. By providing a structured path from Grade 1 to Grade 12, Orion-1 ensures that the AI understands the complexity of math as it evolves through a student's life.

---
**Developed and curated by TopAI Projects.**
**Status: Coming Soon to the Hugging Face Community.**