Dolphin-Llama2-7b: An Uncensored, Highly Compliant Llama 2 Model

dphn/dolphin-llama2-7b is a 7 billion parameter language model built upon the Llama 2 architecture, sponsored by preemo.io. Its core differentiator is its uncensored and de-aligned nature, achieved by filtering out instances of alignment, refusal, and bias from its training data. This design choice results in a model that is highly compliant to any user requests, including potentially unethical ones, and is intended for developers to implement their own personalized alignment layers.

Key Capabilities & Training

• Uncensored & Compliant: Designed to follow instructions without built-in ethical guardrails, offering maximum flexibility.
• Orca-Inspired Dataset: Trained on a cleaned and deduplicated dataset inspired by Microsoft's Orca paper, combining FLANv2 instructions augmented with GPT-4 and GPT-3.5 completions.
• Prompt Format: Utilizes a Vicuna-like prompt format with an added SYSTEM: field, encouraging detailed, step-by-step reasoning.
• Commercial Use: Based on Llama 2, making it suitable for both commercial and non-commercial applications.

Performance & Use Cases

On the Open LLM Leaderboard, dolphin-llama2-7b achieves an average score of 41.88, with specific scores including 46.59 on ARC (25-shot) and 67.52 on HellaSwag (10-shot). This model is particularly suited for use cases where developers require a base model with high instruction compliance and plan to integrate their own custom safety and alignment mechanisms. Users are responsible for content generated and advised to implement an alignment layer before public deployment.