AI Anytime Community
Ablate: Directional Ablation Toolkit for Open LLMs
In short
Ablate is a research toolkit from AI Anytime for mechanistic interpretability and AI safety. It finds the linear direction in a transformer's residual stream that mediates a behaviour — by default refusal — and removes it at runtime or permanently, using a KL-divergence-guided search to balance the change against capability damage. It was demoed live at AI Tinkerers Dubai in October 2026.
Find the direction
Ablate locates the residual-stream direction that mediates a behaviour, using short trigger prompts that exist only to find it.
Remove it, carefully
Ablate at runtime with forward hooks, or permanently by orthogonalising weights; an automatic search tunes which direction, how strongly and which layers.
Measure the cost
An LLM-judge harness and KL divergence quantify both the behaviour change and the damage to the model's other capabilities.
Built for small models
GPT-2, SmolLM2, TinyLlama and Qwen2.5 0.5B–1.5B — models you can study on a laptop or a free Colab GPU.
Key facts
- Vertical
- AI Anytime Community
- For
- Researchers and engineers studying how safety behaviour is represented in small open models
- Price
- Source available on GitHub · pip install ablate-llm
What is Ablate?
Ablate is AI Anytime's toolkit for directional ablation (abliteration): it finds the residual-stream direction that mediates refusal in small open language models and removes it, for interpretability and safety research.
Where was Ablate presented?
Ablate was demoed live at AI Tinkerers Dubai in October 2026.
Is Ablate open source?
The source is publicly available on GitHub for viewing and evaluation, and the package installs with pip install ablate-llm. Other uses require written permission under its licence.
Is Ablate meant for misuse?
No. It is dual-use research tooling for studying how safety behaviour is represented and how robust it is. Use it only on models and in contexts you are authorised to.
More from AI Anytime Community
AI Anytime Academy
University students, software engineers, forward deployed engineers, architects and anyone building practical AI skills
Awesome MCP Servers
Developers and teams building AI agents with Claude, Cursor, VS Code and other MCP clients, and anyone who maintains an MCP server
Jev Router
AI engineers and platform teams running LLM applications, agents and gateways
AI Engineer Roadmap 2027
Students, working professionals and engineers moving into AI
Next step
Let's build AI capability together.
Companies, campuses, partners and sponsors — tell us what you're trying to do and the right AI Anytime team will get back within one working day.