DeepScaleR-1.5B-Preview on Ollama

DeepScaleR-1.5B-Preview is a fine-tuned language model based on DeepSeek-R1-Distilled-Qwen-1.5B, optimized for mathematical reasoning and long-context understanding. This guide provides instructions to run the model locally using Ollama.

Features

Optimized for competition-level math tasks
Supports 8K, 16K, and 24K context lengths
Trained using Group Relative Policy Optimization (GRPO)
Achieves 43.1% Pass@1 accuracy on AIME 2024
Outperforms OpenAI’s O1-Preview despite being only 1.5B parameters

Installation

1. Install Ollama

If you haven’t installed Ollama, follow these steps:

For macOS & Linux:

Run the following command in your terminal:

curl -fsSL https://ollama.com/install.sh | sh

For Windows (WSL required):

wsl --install
curl -fsSL https://ollama.com/install.sh | sh

For more details, check the official Ollama installation guide:
https://ollama.com/download

Running the Model

2. Pull the Model

Once Ollama is installed, download DeepScaleR-1.5B-Preview:

ollama pull yasserrmd/DeepScaleR-1.5B-Preview

3. Run the Model

To start generating responses:

ollama run yasserrmd/DeepScaleR-1.5B-Preview

For interactive chat mode:

ollama chat yasserrmd/DeepScaleR-1.5B-Preview

Customizing the Model

Using a Custom Modelfile

You can modify the model behavior using a Modelfile.

Create a file named Modelfile and add the following content:

FROM yasserrmd/DeepScaleR-1.5B-Preview

PARAMETER temperature 0.7
PARAMETER top_p 0.9

SYSTEM "You are an expert math problem solver. Provide detailed and accurate solutions."

Build and run your customized version:

ollama create my-deepscaler -f Modelfile
ollama run my-deepscaler

Performance Optimization

For improved inference speed: - Run Ollama on a GPU-enabled system - Use quantized versions (e.g., fp16, int4) - Deploy on high-memory cloud instances (24GB RAM or more)

License

This model is released under the MIT License, promoting open-source AI development.

Resources

DeepScaleR-1.5B on Hugging Face:
https://huggingface.co/agentica-org/DeepScaleR-1.5B-Preview
Ollama Official Documentation:
https://ollama.com/docs
GitHub Repository:
https://github.com/agentica-project/deepscaler

Acknowledgments

This model is built on DeepSeek-R1-Distilled-Qwen-1.5B, trained using GRPO at the Berkeley Sky Computing Lab and Berkeley AI Research.

Models

Readme