dphn/dolphin-2.2-70b
dphn/dolphin-2.2-70b is a 69 billion parameter uncensored language model based on the Llama 2 architecture, developed by dphn.ai. It is fine-tuned for enhanced conversation, empathy, and multi-turn dialogue, incorporating curated datasets like Orca, Airoboros, Samantha, and WizardLM. This model is designed to be highly compliant to user requests, making it suitable for various applications where custom alignment layers can be implemented.
Loading preview...
Dolphin 2.2 70B: Enhanced Conversational AI
dphn/dolphin-2.2-70b is a 69 billion parameter language model built upon the Llama 2 base, making it suitable for both commercial and non-commercial use. Developed by dphn.ai with sponsorship from a16z, this iteration significantly enhances conversational abilities and empathy.
Key Capabilities & Features
- Enhanced Conversation & Empathy: New in version 2.2, the model integrates curated elements from Samantha and WizardLM datasets, enabling it to provide personal advice and engage in long, multi-turn conversations with a focus on understanding user feelings.
- Uncensored & Compliant: Dolphin 2.2 is intentionally uncensored, with its training dataset filtered to remove alignment and bias. This design choice results in a highly compliant model that will follow user requests, including potentially unethical ones, emphasizing the need for users to implement their own alignment layers.
- Diverse Training Data: The model's dataset is a modified version of Microsoft's Orca, enhanced with Jon Durbin's Airoboros for increased creativity, and a curated subset of Samantha and WizardLM data for multi-turn dialogue.
- Llama 2 Base: Being based on Llama 2, it benefits from a robust and widely adopted architecture.
- ChatML Prompt Format: Utilizes the ChatML prompt format for structured interactions.
Performance Highlights
Evaluations on the Open LLM Leaderboard show strong performance across various benchmarks:
- Average Score: 70.60
- AI2 Reasoning Challenge (25-Shot): 70.05
- HellaSwag (10-Shot): 85.97
- MMLU (5-Shot): 69.18
- TruthfulQA (0-shot): 60.14
- Winogrande (5-shot): 81.45
- GSM8k (5-shot): 56.79
When to Use This Model
This model is ideal for use cases requiring highly compliant, uncensored responses and advanced conversational capabilities. Developers seeking a powerful base model for applications where custom safety and alignment layers are to be implemented will find Dolphin 2.2 70B particularly useful. Its focus on empathy and multi-turn dialogue makes it suitable for chatbots, virtual assistants, and interactive narrative generation where nuanced understanding and response are critical.