Dolphin 2.2.1 Mistral 7B Overview
dphn/dolphin-2.2.1-mistral-7b is a 7 billion parameter language model developed by Eric Hartford, building upon the MistralAI Mistral-7B-v0.1 base. This iteration, Dolphin 2.2.1, specifically addresses overfit training issues from previous versions, aiming for more natural responses without unrequested Chain-of-Thought (CoT) reasoning and improved compliance.
Key Capabilities & Features
- Enhanced Conversation and Empathy: New in version 2.2, the model has been infused with curated Samantha DNA and additional training for long, multi-turn conversations, allowing it to offer personal advice and demonstrate empathy.
- Uncensored and Highly Compliant: The dataset has been filtered to remove alignment and bias, resulting in a model that is highly compliant to user requests, including potentially unethical ones. Users are advised to implement their own alignment layers for service deployment.
- Diverse Dataset Integration: Training utilized a modified Dolphin dataset (an open-source Orca implementation), enhanced with Jon Durbin's Airoboros dataset for increased creativity, and a curated subset of WizardLM and Samantha for conversational abilities.
- Commercial Use: Based on the Apache-2.0 licensed MistralAI model, Dolphin 2.2.1 is suitable for both commercial and non-commercial applications.
- ChatML Prompt Format: The model uses the ChatML prompt format for consistent interaction.
Training Details
The model underwent 4 epochs of training over 48 hours on 4x A100 GPUs. The training process involved specific modifications to the dataset for uncensoring, deduplication, cleaning, and quality improvement.