README.md

---
base_model:
- deepseek-ai/DeepSeek-R1-Distill-Qwen-7B
datasets:
- VanishD/DualDistill
language:
- en
license: mit
pipeline_tag: text-generation
library_name: transformers
---

# Agentic-R1: Distilled Dual-Strategy Reasoning

This repository hosts the **Agentic-R1** model, an implementation of the paper [**Agentic-R1: Distilled Dual-Strategy Reasoning**](https://huggingface.co/papers/2507.05707).

**Code**: https://github.com/StigLidu/DualDistill

## Abstract

Current long chain-of-thought (long-CoT) models excel at mathematical reasoning but rely on slow and error-prone natural language traces. Tool-augmented agents address arithmetic via code execution, but often falter on complex logical tasks. We introduce a fine-tuning framework, DualDistill, that distills complementary reasoning strategies from multiple teachers into a unified student model. Using this approach, we train Agentic-R1, which dynamically selects the optimal strategy for each query, invoking tools for arithmetic and algorithmic problems, and using text-based reasoning for abstract ones. Our method improves accuracy across a range of tasks, including both computation-intensive and standard benchmarks, demonstrating the effectiveness of multi-strategy distillation in achieving robust and efficient reasoning.

## Key Features

-   **Efficient Training**: Integrates tool use into long-chain-of-thought (CoT) reasoning using only 4 × A6000 GPUs
-   **Unified Reasoning**: Fuses heterogeneous reasoning traces from multiple teacher models into a single student model

<div align="center">
  <img src="https://github.com/StigLidu/DualDistill/raw/main/fig/overview.png" alt="Overview of DualDistill" width="500">
  <p><em>Overview of DualDistill methodology</em></p>
</div>

## Datasets

| Dataset       | Description                                   | Link                                                 |
| :------------ | :-------------------------------------------- | :--------------------------------------------------- |
| **Training Set** | Complete training dataset with teacher trajectories | [🤗 HuggingFace](https://huggingface.co/datasets/VanishD/DualDistill) |
| **Test Set**  | Evaluation benchmarks                         | `dataset/test/`                                      |

## Results

<div align="center">
  <img src="https://github.com/StigLidu/DualDistill/raw/main/fig/result.png" alt="Performance comparison of Agentic-R1 models" width="700">
</div>

-   **Agentic-R1** demonstrates significant performance gains on **DeepMath-L** and **Combinatorics300**, where both complex reasoning and tool use are crucial for success.
-   **Agentic-R1-SD** (Self-Distilled) further enhances performance through our self-distillation approach, consistently outperforming baseline models across nearly all evaluation tasks.

## Quick Start

### Installation

1.  **Clone the repository**:
    ```bash
    git clone https://github.com/StigLidu/DualDistill.git
    cd DualDistill
    ```

2.  **Create environment** (optional but recommended):
    ```bash
    conda create -n dualdistill python=3.11
    conda activate dualdistill
    ```

3.  **Install dependencies**:
    ```bash
    pip install -r requirements.txt
    pip install flash-attn --no-build-isolation
    ```

### Sample Usage

Here's how to perform inference with the `Agentic-R1` model using the Hugging Face `transformers` library:

```python
import torch
from transformers import AutoTokenizer, AutoModelForCausalLM

model_id = "VanishD/Agentic-R1" # Or "VanishD/Agentic-R1-SD" for the self-distilled version

tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype=torch.bfloat16, # Use bfloat16 for better performance and memory if supported
    device_map="auto",
    trust_remote_code=True
).eval() # Set model to evaluation mode

# Prepare a simple user message
messages = [{"role": "user", "content": "What is 123 + 456?"}]

# Apply the chat template to format the prompt correctly for the model
# The `add_generation_prompt=True` adds the Assistant token to prompt the model for its response.
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)

# Encode the prompt
input_ids = tokenizer(prompt, return_tensors="pt").input_ids.to(model.device)

# Generate response
output_ids = model.generate(
    input_ids,
    max_new_tokens=256,
    do_sample=True,
    temperature=0.7,
    top_p=0.95,
    eos_token_id=tokenizer.eos_token_id,
    pad_token_id=tokenizer.pad_token_id, # Often EOS token is used as PAD token for LLMs
)

# Decode and print the generated text, excluding the input prompt
generated_text = tokenizer.decode(output_ids[0][input_ids.shape[1]:], skip_special_tokens=True).strip()
print(f"Generated Text:
{generated_text}")
```

## ⚠️ Important Notes

-   **Code Execution Safety**: The evaluation scripts execute model-generated code locally. Only use trusted models before execution.
-   **Inference Config**: If you are using vLLM (a recent version) and encounter an error regarding the maximum context length. You may need to modify the `model_max_length` in `tokenizer_config.json`.
-   **Self-Distillation Warning**: The self-distillation step requires sampling many trajectories and can be time-consuming.

## License

This project is licensed under the MIT License - see the [LICENSE](https://github.com/StigLidu/DualDistill/blob/main/LICENSE) file for details.

## Acknowledgments

We thank the following open-source projects for their foundational contributions:

-   [OpenHands](https://github.com/All-Hands-AI/OpenHands) - Agent framework
-   [DeepMath-103K](https://huggingface.co/datasets/zwhe99/DeepMath-103K) - Mathematical reasoning dataset
-   [vLLM](https://github.com/vllm-project/vllm) - High-performance inference engine

## Contact

For questions or support, please contact:

-   **Weihua Du**: [[email protected]](mailto:[email protected])

## Citation

If you find our work useful, please consider citing:

```bibtex
@article{du2025agentic,
  title={Agentic-R1: Distilled Dual-Strategy Reasoning},
  author={Du, Weihua and Aggarwal, Pranjal and Welleck, Sean and Yang, Yiming},
  journal={arXiv preprint arXiv:2507.05707},
  year={2025}
}
```