Open-source research code and public model hubs for open-weight LLMs with fewer refusals ("abliteration"). For research, creative writing and local use. You are responsible for how you use them.
A training-free tool that changes model behaviour, for example removing censorship and reducing sycophancy.
A Python library for ablating features in LLMs, built on TransformerLens.
Refusal-direction removal using plain Hugging Face Transformers.
Code for the paper "Refusal in Language Models Is Mediated by a Single Direction".
A community repo of model-unlocking experiments. Read the code before you run it.
A widely shared Hugging Face blog post that explains the technique step by step.
Publishes many "abliterated" versions of popular open-weight models, including GGUF builds.
The author of the abliteration guide. Publishes abliterated and merged models.
The team behind the Dolphin models, with abliterated and "laser" variants.
Example system prompts published alongside the Dolphin models.
Hermes-series instruction-tuned open models that are known for following system prompts closely.
Runs GGUF and open-weight models locally.
The C/C++ engine for GGUF models.
Runs GGUF models in one file, popular for creative writing.
Links point to the original repos and Hugging Face pages. This site hosts no model files. Follow each model's license and your local laws.