Hey Product Hunt! 👋 I'm Sonu, founder of AI Anytime. I built Ablate to make research into LLM refusal behavior more systematic and measurable. Transformer models encode complex behaviors in their internal representations. Ablate gives researchers a practical workflow to investigate those representations and test what happens when specific directions are modified. 🔬 What you can do with Ablate: 1. Extract candidate refusal directions from model activations 2. Experiment with runtime and permanent weight ablation 3. Automatically tune interventions using KL divergence 4. Compare baseline and modified models using evaluation benchmarks and LLM judges 5. Explore the workflow through a local browser UI or Python API One lesson from building this: removing obvious refusal phrases is not the same as demonstrating genuine behavioral change. Evaluation quality matters. I'd love feedback from researchers and ML engineers: What would make this toolkit more useful for studying model behavior, robustness, and safety? 🔗 GitHub: https://github.com/AIAnytime/ablate

Ablate

Understand and measure refusal behavior in open-source LLMs.

Understand and measure refusal behavior in open-source LLMs.