README.md

---
library_name: transformers
license: apache-2.0
license_link: https://huggingface.co/Intel/hebrew-math-tutor-v1/blob/main/LICENSE
pipeline_tag: text-generation
language:
- he
- en
tags:
- mathematics
- education
- hebrew
- reasoning
- math
- tutoring
base_model:
- Qwen/Qwen3-4B-Thinking-2507
---

# Hebrew Math Tutor

<p align="center">
  <img src="https://cdn-uploads.huggingface.co/production/uploads/62d93cd728f9c86a4031562e/YxvxPWRpINziJaAftl4XE.png" width="600"/>
</p>

**Hebrew Math Tutor** is a specialized mathematical reasoning model that provides step-by-step solutions to math problems in Hebrew. Built on Qwen3-4B-Thinking-2507, this model bridges the gap between advanced AI mathematical capabilities and Hebrew-language education.


- 🎯 **Model ID**: `Intel/hebrew-math-tutor-v1`
- 🏗️ **Base Model**: [Qwen3-4B-Thinking-2507](https://huggingface.co/Qwen/Qwen3-4B-Thinking-2507)
- 🏛️ **Architecture**: Decoder-only causal language model (~4B parameters)
- 🗣️ **Primary Language**: Hebrew (retains multilingual capabilities)
- 📄 **License**: Apache-2.0

## Model Description

Hebrew Math Tutor is a supervised fine-tune of Qwen3-4B-Thinking, specifically optimized to:

- **Provide detailed mathematical reasoning in Hebrew** with clear step-by-step explanations
- **Maintain mathematical accuracy** while adapting to Hebrew language patterns  
- **Preserve multilingual capabilities** for cross-language mathematical workflows
- **Support educational applications** with natural Hebrew mathematical discourse

The model excels at translating complex mathematical concepts into clear, pedagogically sound Hebrew explanations while maintaining the computational precision of its base model.

## Intended Use Cases

### ✅ **Primary Applications**

- **Educational Technology**: Hebrew-language math tutoring systems and learning platforms.
- **Research Tools**: Mathematical reasoning research in Hebrew educational contexts.  
- **Prototype Development**: Building Hebrew-first educational AI applications.
- **Accessibility**: Providing advanced math AI assistance to Hebrew-speaking communities.

### ✅ **Secondary Applications**

- Multilingual educational workflows requiring Hebrew mathematical explanations.
- Cross-cultural mathematics education research.
- Hebrew mathematical content generation for educational materials.

### ❌ **Not Intended For**

- **High-stakes assessments**: Medical, legal, or financial decision-making.
- **Unsupervised grading**: Certification or evaluation without human verification.
- **Production systems**: Critical applications without proper validation and oversight.

## Model Details

| **Specification**     | **Details**                                      |
|-----------------------|--------------------------------------------------|
| **Architecture**      | Decoder-only transformer (causal language model) |
| **Parameters**        | ~4 billion                                       |
| **Context Length**    | Inherited from Qwen3-4B-Thinking-2507            |
| **Tokenizer**         | Qwen3-compatible tokenizer with Hebrew support   |
| **Training Type**     | Supervised Fine-Tuning (Hebrew SFT)              |
| **Base Model**        | Qwen3-4B-Thinking-2507                           |
| **Fine-tuning Focus** | Mathematical reasoning in Hebrew                 |

## Training Details

### **Dataset**

- **Source**: ~10,000 selected problems from [OpenMathReasoning](https://huggingface.co/datasets/nvidia/OpenMathReasoning).
- **Translation Approach**: Automated high-quality translation using internal LLMs.
- **Language Adaptation**: Questions and final answers translated to Hebrew; reasoning chains preserved.
- **Mathematical Notation**: Equations and formal math notation kept intact.
- **Internal Reasoning**: Model's `<think>...</think>` blocks intentionally remain in English (representing internal reasoning processes).

### **Training Configuration**

- **Method**: Supervised Fine-Tuning (Hebrew SFT)
- **Epochs**: 3
- **Learning Rate**: 5e-6
- **Warmup**: 0.1
- **Scheduler**: Cosine learning rate decay
- **Objective**: Maintain mathematical accuracy while adapting output to Hebrew

## Performance Evaluation

We evaluated Hebrew Math Tutor on three challenging mathematical benchmarks: **MATH500**, **AIME24**, and **AIME25**.

### **Evaluation Metrics**

- **pass@16**: Percentage of problems where at least one of 16 generated samples is correct.
- **maj@16**: Majority-vote accuracy across 16 samples.
- **Hebrew Answers**: Percentage of responses generated in Hebrew.

### **Hebrew Evaluation Results**

| Dataset     | Metric         | Base Model | Hebrew Math Tutor | Improvement |
|-------------|----------------|------------|-------------------|-------------|
| **MATH500** | pass@16        | 93%        | **95%**           | +2%         |
|             | maj@16         | 88%        | **90%**           | +2%         |
|             | Hebrew Answers | 75%        | **100%**          | +25%        |
| **AIME24**  | pass@16        | 76.7%      | **80%**           | +3.3%       |
|             | maj@16         | 76.7%      | **76.7%**         | No change   |
|             | Hebrew Answers | 35.2%      | **96.7%**         | +61.5%      |
| **AIME25**  | pass@16        | 80%        | **83.3%**         | +3.3%       |
|             | maj@16         | 70%        | **60%**           | -10%        |
|             | Hebrew Answers | 36%        | **95.2%**         | +59.2%      |

### **English/Original Language Results**

| Dataset     | Metric  | Base Model | Hebrew Math Tutor | Change    |
|-------------|---------|------------|-------------------|-----------|
| **MATH500** | pass@16 | 99%        | **98%**           | -1%       |
|             | maj@16  | 98%        | **98%**           | No change |
| **AIME24**  | pass@16 | 93.3%      | **90%**           | -3.3%     |
|             | maj@16  | 86.7%      | **86.7%**         | No change |
| **AIME25**  | pass@16 | 83.3%      | **90%**           | +6.7%     |
|             | maj@16  | 73%        | **80%**           | +7%       |

### **Key Findings**

🎯 **Dramatic Language Improvement**: Hebrew answer generation increased by 25-61.5% across all benchmarks, reaching 95-100% Hebrew output.

📈 **Maintained Technical Performance**: Consistent improvements in pass@16 on Hebrew evaluations while preserving competitive English performance.

🔍 **Mixed Majority Vote Results**: Strong performance on MATH500, stable on AIME24, with one notable decrease on AIME25 requiring further investigation.

✅ **Preserved Core Capabilities**: The fine-tuning successfully adapted language output without sacrificing fundamental mathematical reasoning abilities.

## Usage

### **Quick Start**

```python
from transformers import pipeline

model = "Intel/hebrew-math-tutor-v1"
pipe = pipeline("text-generation", model)

messages = [
    {
        "role": "system",
        "content": """You are a helpful AI assistant specialized in mathematics and problem-solving who can answer math questions with the correct answer.
Answer shortly, not more than 500 tokens, but outline the process step by step.
Answer ONLY in Hebrew!""",
    },
    {"role": "user", "content": "מהו סכום הסדרה הבאה:  1 + 1/2 + 1/4 + 1/8 + ..."},
]

out = pipe(
    messages,
    return_full_text=False,
    max_new_tokens=1024,
    temperature=0.6,
    top_p=0.95,
    top_k=20,
)
print(out[0]["generated_text"])
```

### **Recommended Parameters**

- **Temperature**: 0.6 (balanced creativity and accuracy)
- **Top-p**: 0.95 (diverse but focused sampling)
- **Top-k**: 20 (controlled vocabulary selection)
- **Max tokens**: 500-1024 (sufficient for detailed explanations)

### **Best Practices**

- **Request explicit structure**: Ask for step-by-step reasoning and clearly marked final answers.
- **Use Hebrew formatting cues**: Include phrases like "תשובה סופית:" or request `\boxed{}` formatting.
- **Specify language**: Explicitly request Hebrew-only responses for consistent output.
- **Verify solutions**: Always validate mathematical results, especially in educational contexts.

## Demo Interface

<p align="center">
    <img src="https://cdn-uploads.huggingface.co/production/uploads/62d93cd728f9c86a4031562e/tbOIu47QLmja_z-Ce20a2.png" width="600"/>
    <br>
    <em>Example Streamlit interface showing Hebrew Math Tutor providing step-by-step reasoning. The detailed reasoning can be collapsed for cleaner presentation.</em>
</p>

## Limitations & Considerations

### **Technical Limitations**

- **Potential errors**: May produce incorrect solutions or mathematical hallucinations.
- **Language mixing**: Occasional mixing of Hebrew and English or inconsistent number formatting.
- **Training biases**: May reflect biases present in the original training datasets.
- **Internal reasoning**: `<think>...</think>` blocks remain in English due to training scope.

### **Usage Recommendations**

- **Human verification required**: Always validate outputs before use in educational settings
- **Not a replacement for educators**: Designed as an assistive tool, not a substitute for qualified instruction.
- **Appropriate context**: Best suited for educational prototyping and research applications.

## Ethical Guidelines

### **Responsible Deployment**

- Include clear disclaimers about AI-generated content in user-facing applications.
- Implement human oversight for any educational or assessment applications.
- Ensure compliance with relevant privacy laws when collecting user data.
- Provide transparency about model capabilities and limitations.

### **Educational Impact**

- Designed to enhance, not replace, human mathematical instruction.
- Intended to increase accessibility of advanced math AI for Hebrew speakers.
- Should be used as part of comprehensive educational approaches with human guidance.

## Technical Details

### **Evaluation Methodology**

- **Correctness verification**: Solutions validated using Math-verify framework.
- **Statistical significance**: Results based on 16 samples per problem for robust evaluation.
- **Language detection**: Automated classification of response language for Hebrew Answers metric.
- **Benchmark diversity**: Evaluation across competition mathematics (AIME) and curriculum problems (MATH500).

### **Reproducibility**

- All evaluation protocols follow standard mathematical reasoning assessment practices.
- Sampling parameters and evaluation metrics clearly documented.
- Training configuration and hyperparameters provided for reproduction.

## Attribution & Licensing

- **Model License**: [Apache-2.0](https://huggingface.co/Intel/hebrew-math-tutor-v1/blob/main/LICENSE)
- **Base Model**: [Qwen3-4B-Thinking-2507](https://huggingface.co/Qwen/Qwen3-4B-Thinking-2507) (Alibaba)
- **Training Dataset**: [OpenMathReasoning](https://huggingface.co/datasets/nvidia/OpenMathReasoning) (NVIDIA)
- **Development**: Intel Labs

## Citation

If you use Hebrew Math Tutor in your research or applications, please cite:

```bibtex
@misc{hebrew-math-tutor-v1,
  title={Hebrew Math Tutor: A Hebrew-focused Mathematical Reasoning Model},
  author={Intel AI},
  year={2025},
  url={https://huggingface.co/Intel/hebrew-math-tutor-v1},
  note={Fine-tuned from Qwen3-4B-Thinking-2507}
}
```

## Community & Support

- **Blog Post**: [more details in the blog](https://huggingface.co/blog/danf/hebrew-math-tutor).
- **Model Repository**: [https://huggingface.co/Intel/hebrew-math-tutor-v1](https://huggingface.co/Intel/hebrew-math-tutor-v1)
- **Issues & Feedback**: Use the Hugging Face repository issues for bug reports and feature requests.
- **Community Discussions**: Join conversations in the repository discussions tab.

## Changelog

- **v1.0** — Initial public release with Hebrew mathematical reasoning capabilities.

---

*Hebrew Math Tutor represents a step forward in making advanced mathematical AI accessible across languages. We encourage responsible use and welcome community feedback to improve multilingual mathematical reasoning capabilities.*