Hey everyone — I built a Space that takes a different approach to local LLM customization and wanted to share it here.
What it does: Instead of tweaking system prompts to get a “more creative” or “more analytical” model, LoRA Lens Behavior Studio applies behavior-level adapters extracted via contrast pairs and SVD decomposition — then lets you dial the intensity up or down with a signed alpha slider in real time.
The core pipeline:
-
Contrast pair extraction (target behavior vs. baseline)
-
Delta compression into LoRA adapters via SVD (rank 32–64)
-
Live
AlphaControllerfor non-destructive steering — nothing gets baked permanently
Demo: https://huggingface.co/spaces/Intuitivation/LoRALens-Demo
Certified across Mistral-7B, Llama-3-8B, and Qwen2.5 series. 36 behavior packs across Communication, Reasoning, Creative, Personality, Domain-Specific, and Experimental categories.
The thing that surprised people most in testing: dragging the slider from −1.0 to +1.0 on something like the Confident Clarity pack produces a noticeably different feel in the outputs — not just word choice, but sentence structure and hedging behavior. It’s not a prompt effect.
Would love feedback from anyone who’s worked with activation-level steering or has thoughts on the SVD rank tradeoffs. Happy to answer questions about the architecture.