dphn/dolphin-2.1-mistral-7b

Overview

dphn/dolphin-2.1-mistral-7b is a 7 billion parameter language model developed by Eric Hartford, built upon the MistralAI architecture. It is notable for its uncensored nature and high compliance, achieved by filtering the training dataset to remove alignment and bias. The model is released under an Apache-2.0 license, permitting both commercial and non-commercial applications.

Key Capabilities

• Uncensored and Highly Compliant: Designed to be highly responsive to user requests, including potentially unethical ones, requiring users to implement their own alignment layers for responsible deployment.
• Dataset Origin: Trained on a modified version of the Dolphin dataset, which is an open-source implementation of Microsoft's Orca, enhanced with Jon Durbin's Airoboros dataset for increased creativity.
• Training Details: The model underwent 4 epochs of training over 48 hours using 4x A100 GPUs.
• Prompt Format: Utilizes the ChatML prompt format, consistent with future releases from the developer.

Performance

Evaluated on the Open LLM Leaderboard, dolphin-2.1-mistral-7b achieved an average score of 53.47. Specific benchmark results include:

• ARC (25-shot): 64.42
• HellaSwag (10-shot): 84.92
• MMLU (5-shot): 63.32
• TruthfulQA (0-shot): 55.56
• Winogrande (5-shot): 77.74
• GSM8K (5-shot): 20.77
• DROP (3-shot): 7.56

Good For

• Applications requiring a highly compliant and uncensored language model.
• Developers who need a base model to implement custom alignment and safety layers.
• Use cases benefiting from a model trained on a blend of Orca-like explanation traces and creative datasets.