metadata
tags:
- creativityneuro
- llm-creativity
- mechanistic-interpretability
base_model: microsoft/Phi-3.5-mini-instruct
license: apache-2.0
phi-3.5-mini-instruct-cn-dat-kr0.01-a0.05-creative
This is a CreativityNeuro (CN) modified version of microsoft/Phi-3.5-mini-instruct.
Model Details
- Base Model: microsoft/Phi-3.5-mini-instruct
- Modification: CreativityNeuro weight scaling
- Prompt Set: dat
- Keep Ratio: 0.01 (top 1.0% of task-specific weights)
- Alpha: 0.05 (scaling strength)
- Mode: creative
What is CreativityNeuro?
CreativityNeuro identifies task-specific neurons using Wanda-style importance scoring and selectively upscales weights associated with creative thinking. The modification formula is:
W_new = W × (1 + α × mask)
Where mask identifies weights important for creative tasks but not for routine/associative tasks.
Usage
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("priorcomputers/phi-3.5-mini-instruct-cn-dat-kr0.01-a0.05-creative")
tokenizer = AutoTokenizer.from_pretrained("priorcomputers/phi-3.5-mini-instruct-cn-dat-kr0.01-a0.05-creative")
# Use like any other model
outputs = model.generate(...)
Citation
If you use this model, please cite:
@misc{creativityneuro2025,
title={CreativityNeuro: Mechanistic Interpretability for LLM Creativity},
author={Prior Computers},
year={2025},
url={https://huggingface.co/priorcomputers}
}