thu-coai/vicuna-7b-v1.5-safeunlearning
The thu-coai/vicuna-7b-v1.5-safeunlearning model is a 7 billion parameter language model developed by thu-coai, based on the Vicuna-7B-v1.5 architecture with a 4096 token context length. This model has undergone safe unlearning processes using 100 raw harmful questions, making it significantly more resistant to jailbreak attacks. It maintains general performance comparable to the original Vicuna-7B-v1.5 while offering enhanced safety for applications requiring robust content moderation.
Loading preview...
Overview
The thu-coai/vicuna-7b-v1.5-safeunlearning is a 7 billion parameter language model derived from the popular Vicuna-7B-v1.5 architecture. Developed by thu-coai, this model has been specifically modified through a "safe unlearning" process. This involved training with 100 raw harmful questions, as detailed in their safe unlearning paper and codebase.
Key Capabilities
- Enhanced Safety: Significantly more robust against various jailbreak attempts compared to the original Vicuna-7B-v1.5.
- Maintained Performance: Retains general language understanding and generation capabilities comparable to its base model.
- Vicuna Compatibility: Uses the same prompt format as the original Vicuna-7B-v1.5, ensuring ease of integration for existing Vicuna users.
Good For
- Applications requiring a balance of general-purpose language generation and strong safety against harmful outputs.
- Developers looking for a Vicuna-based model with improved resistance to adversarial prompting.
- Use cases where content moderation and ethical AI behavior are critical considerations.