rainerwinkler123456789/Llama-3.3-70B-Instruct-abliterated
The rainerwinkler123456789/Llama-3.3-70B-Instruct-abliterated model is a 70 billion parameter instruction-tuned causal language model, derived from Meta's Llama-3.3-70B-Instruct. Developed by huihui-ai, this model has been modified using an 'abliteration' technique to remove refusal behaviors, offering an uncensored response capability. It is designed for use cases requiring direct answers without built-in content moderation, supporting a 32768 token context length.
Loading preview...
Model Overview
This model, rainerwinkler123456789/Llama-3.3-70B-Instruct-abliterated, is a 70 billion parameter instruction-tuned large language model. It is a modified version of the original meta-llama/Llama-3.3-70B-Instruct model, developed by huihui-ai.
Key Differentiator: Abliteration for Uncensored Responses
The primary distinction of this model is its application of an "abliteration" technique, a proof-of-concept method to remove refusal behaviors from the LLM. This process aims to provide responses without the built-in content moderation or refusal mechanisms often found in instruction-tuned models. The technique is implemented without relying on TransformerLens, as detailed in the remove-refusals-with-transformers project.
Usage and Capabilities
- Uncensored Output: Designed to provide direct answers, even to prompts that might typically trigger refusal responses in other models.
- Instruction Following: Retains the instruction-following capabilities of its base Llama-3.3-70B-Instruct model.
- Context Length: Supports a substantial context window of 32768 tokens.
- Flexibility: Users are encouraged to re-prompt if initial responses are not satisfactory, as the uncensoring process is a proof-of-concept.
Deployment
This model can be easily integrated into applications using the transformers library (version 4.43.0 or newer) for conversational inference. It is also available for use with ollama under the name huihui_ai/llama3.3-abliterated.