saukko/Abliterated-Dolphin3.0-R1-Mistral-24B
The saukko/Abliterated-Dolphin3.0-R1-Mistral-24B is a 24 billion parameter Mistral-based language model, derived from Dolphin3.0-R1-Mistral-24B, that has undergone an 'abliteration' process. This experimental model is intentionally 'dumbed down' and not recommended for meaningful applications, serving primarily as a demonstration of the abliteration technique to remove censorship. It features a 32768 token context length but is noted to have significant limitations compared to its base model.
Loading preview...
Abliterated-Dolphin3.0-R1-Mistral-24B Overview
This model, created by saukko, is an experimental 24 billion parameter Mistral-based language model. It is a modified version of the original Dolphin3.0-R1-Mistral-24B, specifically processed using a technique called "abliteration." The primary purpose of this modification was to explore the effects of abliteration, a method aimed at redirecting refusals and effectively removing censorship from a language model, as inspired by mlabonne's work.
Key Characteristics & Limitations
- Experimental Nature: The model is a result of curiosity regarding the abliteration technique, not intended for practical, meaningful applications.
- "Dumbed Down": It is explicitly stated to be a "dumbed down" version of its base model, as no DPO (Direct Preference Optimization) fine-tuning was performed to mitigate the damage caused by the abliteration process.
- High Risks: The creator advises against using this model for serious tasks, recommending the original Dolphin3.0-R1-Mistral-24B instead due to its significantly increased risks and limitations.
- Context Length: Maintains the 32768 token context length of its base.
Intended Use Case
This model is primarily for demonstrative and exploratory purposes related to the abliteration technique. It is not recommended for general use cases where reliability, accuracy, or safety are important. Users interested in the underlying concepts are encouraged to consult the references provided, particularly mlabonne's article on abliteration.