dphn/dolphin-2.9.2-qwen2-72b
Dolphin 2.9.2 Qwen2 72B is a 72.7 billion parameter instruction-tuned causal language model developed by Eric Hartford, Lucas Atkins, Fernando Fernandes, and Cognitive Computations. Based on the Qwen2-72B architecture, this model is uncensored and designed for high compliance with requests, offering strong instruction following, conversational abilities, coding skills, and initial agentic capabilities with function calling. It is fine-tuned with an 8k sequence length, making it suitable for diverse applications requiring a highly compliant and versatile large language model.
Loading preview...
Dolphin 2.9.2 Qwen2 72B Overview
Dolphin 2.9.2 Qwen2 72B is a powerful 72.7 billion parameter language model developed by Eric Hartford, Lucas Atkins, Fernando Fernandes, and Cognitive Computations. It is built upon the Qwen2-72B base model and has been fine-tuned with a focus on instruction following, conversational fluency, and coding proficiency. A key characteristic of this model is its uncensored nature, designed for high compliance with user requests, including those that might be considered unethical, emphasizing the need for users to implement their own alignment layers.
Key Capabilities
- Instruction Following: Excels at understanding and executing complex instructions.
- Conversational Skills: Capable of engaging in natural and coherent dialogues.
- Coding Abilities: Demonstrates strong performance in various coding tasks.
- Agentic Features: Possesses initial agentic capabilities, including support for function calling.
- High Compliance: Designed to be highly compliant with user prompts, offering flexibility for diverse applications.
Training and Licensing
The model was fine-tuned using a full-weight approach with an 8k sequence length, leveraging parameters selected by Laser Scanner. It utilizes the ChatML prompt template format. Dolphin 2.9.2 is licensed under the tongyi-qianwen license, permitting commercial use in accordance with its terms. The training data included content generated from GPT-4, among other sources.
Performance Highlights
Evaluations on the Open LLM Leaderboard show an average score of 32.00. Specific metrics include:
- IFEval (0-Shot): 40.38
- BBH (3-Shot): 47.70
- MATH Lvl 5 (4-Shot): 21.37
- MMLU-PRO (5-shot): 49.52
Good For
- Developers requiring a highly compliant and uncensored model for research or specific applications.
- Use cases demanding strong instruction following, conversational AI, or code generation.
- Applications where initial agentic capabilities and function calling are beneficial.