Overview

This model, llama-3.1-8b-instruct-cn-dat-kr0.1-a1.0-creative, is an 8 billion parameter instruction-tuned variant of Meta's Llama-3.1-8B-Instruct, developed by Prior Computers. It incorporates a unique modification called CreativityNeuro (CN) weight scaling, specifically designed to enhance creative thinking capabilities.

Key Modifications

• CreativityNeuro (CN) Weight Scaling: This technique identifies and selectively upscales task-specific neurons important for creative tasks, using a Wanda-style importance scoring method.
• Modification Formula: W_new = W × (1 + α × mask), where mask highlights weights crucial for creative tasks, and α (alpha) is the scaling strength, set to 1.0 in this version.
• Keep Ratio: Only the top 10.0% of task-specific weights are targeted for modification.
• Mode: Explicitly set for "creative" tasks.

Use Cases

This model is particularly suited for applications requiring enhanced creative text generation, such as:

• Creative writing and storytelling
• Brainstorming and idea generation
• Content creation where novelty and imaginative output are desired

Citation

If you utilize this model in your work, please consider citing the associated research on CreativityNeuro by Prior Computers.