This project enables users to extend the reasoning capabilities of DeepSeek R1 models, allowing them to "overthink" and produce more thorough responses. It targets researchers and power users seeking deeper insights from LLMs by providing fine-grained control over the generation process and supporting unlimited context length, limited only by VRAM.

How It Works

The core mechanism involves intercepting early model conclusions and replacing them with prompts that encourage further deliberation. This "budget forcing" technique, validated by the independent "s1: Simple test-time scaling" paper, allows for controlled extension of the model's thinking process until a user-defined threshold is met. It leverages unsloth-optimized models for enhanced performance and VRAM efficiency.

Quick Start & Requirements

• Install via pip install -e .
• Requires Python 3.10+ and PyTorch.
• Supports various DeepSeek R1 models (1.5B to 70B parameters), including Qwen and LLaMA architectures.
• Models up to 14B parameters can run on a free Google Colab T4 GPU.
• See unsloth for optimization details.

Highlighted Details

• Forces models to think longer and more thoroughly.
• Customizable reasoning extensions and thinking thresholds.
• Fine-grained control over model parameters (temperature, top-p).
• Visible thinking process with token count tracking.

Maintenance & Community

• Developed by anzorq.
• Credits original idea to vgel's gist.
• Utilizes unsloth for optimization and Gradio for the app interface.

Licensing & Compatibility

• MIT License.
• Permissive license suitable for commercial use and integration into closed-source projects.

Limitations & Caveats

The effectiveness of "overthinking" may vary depending on the specific model and task. The project relies on unsloth optimizations, which might introduce specific dependencies or behaviors.