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 Data

The model was trained on a combination of three datasets:

leonvanbokhorst/friction-disagreement-v2 (8.5% weight)
- Examples of productive disagreement and challenging assumptions
leonvanbokhorst/friction-overthinking-v2 (9.5% weight)
- Examples of deep analytical thinking and self-reflection
leonvanbokhorst/reluctance-v6.1 (82% weight)
- Examples of hesitation and careful consideration

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

Performance Metrics

Training Loss: 1.437 (final)
Best Validation Loss: 1.527 (epoch 3.57)
Memory Usage: 3.813 GB for training (15.9% of GPU memory)

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

Usage

Example usage with the Transformers library:

from transformers import AutoModelForCausalLM, AutoTokenizer

# Load model and tokenizer
model_name = "leonvanbokhorst/deepseek-r1-mixture-of-friction"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name)

# Format input with chat template
prompt = """<|im_start|>system
You are a human-like AI assistant.
<|im_end|>
<|im_start|>user
Why do I keep procrastinating important tasks?
<|im_end|>
<|im_start|>assistant"""

# Generate response
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(
    inputs["input_ids"],
    max_length=512,
    temperature=0.7,
    top_p=0.9
)
response = tokenizer.decode(outputs[0], skip_special_tokens=True)

Training Details

LoRA Configuration

Rank: 16
Alpha: 32
Target Modules:
- q_proj
- k_proj
- v_proj
- o_proj
- gate_proj
- up_proj
- down_proj

Dataset Processing

Examples stacked up to 4096 tokens
90/10 train/validation split
Consistent seed (42) for reproducibility
Token-based sampling for balanced training

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, and Hesitation},
  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