---
license: apache-2.0
library_name: transformers
language:
- en
base_model:
- Qwen/Qwen2.5-14B
pipeline_tag: text-generation
---

# Datarus-R1-14B-preview

<div align="center">
  <img src="https://i.postimg.cc/7hsStNgm/logo-icon-2-1.png" alt="Datarus Logo" width="150"/>
  
  [![Model](https://img.shields.io/badge/Model-Datarus--R1--14B-blue)](https://huggingface.co/DatarusAI/Datarus-R1-14B-preview)
  [![License](https://img.shields.io/badge/License-Apache%202.0-green)](LICENSE)
  [![Website](https://img.shields.io/badge/Website-datarus.ai-orange)](https://datarus.ai)
  [![Demo](https://img.shields.io/badge/Demo-Try%20Now-purple)](https://chat.datarus.ai)
  [![Paper](https://img.shields.io/badge/Paper-arXiv-red)](https://arxiv.org/abs/2508.13382)
</div>

## 🚀 Overview

**Datarus-R1-14B-Preview** is a 14B-parameter open-weights language model fine-tuned from Qwen2.5-14B-Instruct, designed to act as a virtual data analyst and graduate-level problem solver. Unlike traditional models trained on isolated Q&A pairs, Datarus learns from complete analytical trajectories—including reasoning steps, code execution, error traces, self-corrections, and final conclusions—all captured in a ReAct-style notebook format.

### Key Highlights

- **🎯 State-of-the-art efficiency**: Surpasses similar-sized models and competes with 32B+ models while using 18-49% fewer tokens
- **🔄 Dual reasoning interfaces**: Supports both Agentic (ReAct) mode for interactive analysis and Reflection (CoT) mode for concise documentation
- **📊 Superior performance**: Achieves up to 30% higher accuracy on AIME 2024/2025 and LiveCodeBench
- **💡 "AHA-moment" pattern**: Exhibits efficient hypothesis refinement in 1-2 iterations, avoiding circular reasoning loops

## 🔗 Quick Links

- 🌐 **Website**: [https://datarus.ai](https://datarus.ai)
- 💬 **Try the Demo**: [https://chat.datarus.ai](https://chat.datarus.ai)
- 🛠️ **Jupyter Agent**: [GitHub Repository](https://github.com/DatarusAI/Datarus-JupyterAgent)
- 📄 **Paper**: [Datarus-R1: An Adaptive Multi-Step Reasoning LLM](https://arxiv.org/abs/2508.13382)

## 📊 Performance

### Benchmark Results

| Benchmark | Datarus-R1-14B-Preview | QwQ-32B | Phi-4-reasoning | DeepSeek-R1-Distill-14B |
|-----------|----------------|---------|-----------------|-------------------------|
| **LiveCodeBench v6** | 57.7 | 56.6 | 52.6 | 48.6 |
| **AIME 2024** | 70.1 | 76.2 | 74.6* | - |
| **AIME 2025** | 66.2 | 66.2 | 63.1* | - |
| **GPQA Diamond** | 62.1 | 60.1 | 55.0 | 58.6 |

*Reported values from official papers

### Token Efficiency and Performance

<div align="center">
    <img src="https://i.postimg.cc/NMSppNM4/perf-efficiency.png" alt="LCB-Efficiency" width="600"/>
    <img src="https://i.postimg.cc/nV341Ssf/efficiency.png" alt="Efficiency" width="600" />
</div>

## 🎯 Model Card

### Model Details

- **Model Type**: Language Model for Reasoning and Data Analysis
- **Parameters**: 14.8B
- **Training Data**: 144,000 synthetic analytical trajectories across finance, medicine, numerical analysis, and other quantitative domains + A curated collection of reasoning datasets.
- **Language**: English
- **License**: Apache 2.0

### Intended Use

#### Primary Use Cases
- **Data Analysis**: Automated data exploration, statistical analysis, and visualization
- **Mathematical Problem Solving**: Graduate-level mathematics including AIME-level problems
- **Code Generation**: Creating analytical scripts and solving programming challenges
- **Scientific Reasoning**: Complex problem-solving in physics, chemistry, and other sciences
- **Interactive Notebooks**: Building complete analysis notebooks with iterative refinement

### Dual Mode Usage

#### Agentic Mode (for interactive analysis)
- Use `<step>`, `<thought>`, `<action>`, `<action_input>`, `<observation>` tags
- Enables iterative code execution and refinement
- Best for data analysis, simulations, and exploratory tasks

#### Reflection Mode (for documentation)
- Use `<think>` and `<answer>` tags
- Produces compact, self-contained reasoning chains
- Best for mathematical proofs, explanations, and reports

## 📚 Citation

```bibtex
@article{benchaliah2025datarus,
  title={Datarus-R1: An Adaptive Multi-Step Reasoning LLM for Automated Data Analysis},
  author={Ben Chaliah, Ayoub and Dellagi, Hela},
  journal={arXiv preprint arXiv:2508.13382},
  year={2025}
}
```

## 🤝 Contributing

We welcome contributions! Please see our [GitHub repository](https://github.com/DatarusAI/Datarus-JupyterAgent) for:
- Bug reports and feature requests
- Pull requests
- Discussion forums

## 📄 License

This model is released under the Apache 2.0 License.

## 🙏 Acknowledgments

We thank the Qwen team for the excellent base model and the open-source community for their valuable contributions.

## 📧 Contact

- **Email**: ayoub1benchaliah@gmail.com, hela.dellagi@outlook.com
- **Website**: [https://datarus.ai](https://datarus.ai)
- **Demo**: [https://chat.datarus.ai](https://chat.datarus.ai)

---

<div align="center">
  <strong>Experience the future of AI-powered data analysis with Datarus-R1</strong>
  
  [Try Demo](https://chat.datarus.ai) | [View Code](https://github.com/DatarusAI/Datarus-JupyterAgent) | [Read Paper](https://arxiv.org/abs/2508.13382)
</div>

## ⭐ Support

If you find this model and Agent pipeline useful, please consider __Like/Star__! Your support helps us continue improving the project.

Found a bug or have a feature request? Please open an issue on GitHub.

---

<p align="center">Made with ❤️ by the Datarus Team from Paris</p>