Hugging Face Overview
jondurbin/airoboros-13b is an experimental 13 billion parameter LLaMA model, fine-tuned by jondurbin. Its unique characteristic lies in its training data, which was entirely synthetically generated using a 'jailbreak' prompt with OpenAI's GPT-4 and GPT-3.5-turbo. This method aimed to bypass OpenAI's alignment filters, resulting in a broader and potentially 'harmful' range of data and fewer refusals on sensitive topics.
Performance
In internal evaluations using GPT-4 judging on 200 prompts, airoboros-13b achieved a raw score of 17947, translating to a 98.087 GPT-3.5 adjusted score. This places it competitively against models like gpt4-x-alpasta-30b, manticore-13b, vicuna-13b-1.1, and wizard-vicuna-13b-uncensored.
Key Characteristics
- Synthetic Data Training: Utilizes a novel approach of generating training data via 'jailbreak' prompts to explore broader content generation.
- LLaMA Base: Built upon the LLaMA architecture.
- Experimental Nature: The developer explicitly states that this model is not recommended for general use due to output quality and potential for harmful content.
Usage and Licensing
The model is intended for research use only and is subject to a custom license due to its LLaMA base and the use of OpenAI-generated data. Commercial use is explicitly prohibited.