indigoskyai/Qwen2.5-72B-Instruct-abliterated
indigoskyai/Qwen2.5-72B-Instruct-abliterated is a 72.7 billion parameter instruction-tuned causal language model based on Qwen2.5-72B-Instruct, developed by huihui-ai. This model has been specifically modified using an 'abliteration' technique to remove refusal behaviors, making it an uncensored version. It maintains the original 32K context length and is primarily designed for applications requiring a large, powerful language model without built-in content restrictions.
Loading preview...
Overview
This model, indigoskyai/Qwen2.5-72B-Instruct-abliterated, is a 72.7 billion parameter instruction-tuned language model derived from the original Qwen2.5-72B-Instruct. Its key differentiator is the application of an "abliteration" technique, as detailed in the remove-refusals-with-transformers project, to remove inherent refusal mechanisms. This makes it an uncensored variant of the base Qwen model, offering a proof-of-concept for modifying LLM behavior without relying on tools like TransformerLens.
Key Capabilities
- Uncensored Responses: Designed to provide direct answers without built-in refusal behaviors, unlike its base model.
- Large Scale: Benefits from the 72.7 billion parameters of the Qwen2.5-72B-Instruct architecture, suitable for complex tasks.
- Instruction Following: Retains the instruction-following capabilities of the original Qwen2.5-72B-Instruct.
- Standard Integration: Easily loadable and usable with Hugging Face's
transformerslibrary, with provided Python code examples. - Ollama Support: Available for direct use via Ollama, simplifying local deployment.
Good For
- Developers and researchers exploring methods for modifying LLM behavior and removing content filters.
- Applications requiring a powerful, large-scale language model that does not exhibit refusal tendencies.
- Use cases where direct, unfiltered responses are preferred or necessary, provided ethical considerations are managed by the user.
Performance
The model's performance is generally aligned with the base Qwen2.5-72B-Instruct model, with specific evaluations available on the open-llm-leaderboard.