leonvanbokhorst/deepseek-r1-mixture-of-friction

Friction Reasoning Model

This model is fine-tuned to engage in productive disagreement, overthinking, and reluctance. It's based on DeepSeek-R1-Distill-Qwen-7B and trained on a curated dataset of disagreement, overthinking, and reluctance examples.

Model Description

• —Model Architecture: DeepSeek-R1-Distill-Qwen-7B with LoRA adapters
• —Language(s): English
• —License: Apache 2.0
• —Finetuning Approach: Instruction tuning with friction-based reasoning examples

Training Procedure

• —Hardware: NVIDIA RTX 4090 (24GB)
• —Framework: Unsloth + PyTorch
• —Training Time: 35 minutes
• —Epochs: 7 (early convergence around epoch 4)
• —Batch Size: 2 per device (effective batch size 8 with gradient accumulation)
• —Optimization: AdamW 8-bit
• —Learning Rate: 2e-4 with cosine schedule
• —Weight Decay: 0.01
• —Gradient Clipping: 0.5
• —Mixed Precision: bfloat16

Intended Use

This model is designed for:

• —Engaging in productive disagreement
• —Challenging assumptions constructively
• —Providing alternative perspectives
• —Deep analytical thinking
• —Careful consideration of complex issues

Limitations

The model:

• —Is not designed for factual question-answering
• —May sometimes be overly disagreeable
• —Should not be used for medical, legal, or financial advice
• —Works best with reflective or analytical queries
• —May not perform well on objective or factual tasks

Bias and Risks

The model:

• —May exhibit biases present in the training data
• —Could potentially reinforce overthinking in certain situations
• —Might challenge user assumptions in sensitive contexts
• —Should be used with appropriate content warnings

Citation

If you use this model in your research, please cite:

@misc{friction-reasoning-2025,
  author = {Leon van Bokhorst},
  title = {Mixture of Friction: Fine-tuned Language Model for Productive Disagreement, Overthinking, Uncertainty and Reluctance},
  year = {2025},
  publisher = {HuggingFace},
  journal = {HuggingFace Model Hub},
  howpublished = {\url{https://huggingface.co/leonvanbokhorst/deepseek-r1-mixture-of-friction}}
}

Acknowledgments

• —DeepSeek AI for the base model
• —Unsloth team for the optimization toolkit
• —HuggingFace for the model hosting and infrastructure