lactroiii/Dolphin3.0-R1-Mistral-24B
Dolphin 3.0 R1 Mistral 24B is a 24 billion parameter instruct-tuned model from the Dolphin 3.0 Collection, curated and trained by Eric Hartford, Ben Gitter, BlouseJury, and Cognitive Computations. This model is designed as a general-purpose local model, excelling in coding, math, agentic tasks, and function calling, with a 32768 token context length. It has been trained for 3 epochs using 800k reasoning traces from the Dolphin-R1 dataset, focusing on general-purpose reasoning and user steerability.
Loading preview...
Dolphin 3.0 R1 Mistral 24B Overview
Dolphin 3.0 R1 is an instruct-tuned model from the Dolphin series, developed by Eric Hartford, Ben Gitter, BlouseJury, and Cognitive Computations. This 24 billion parameter model with a 32768 token context length is engineered to be a versatile local AI, supporting a wide range of applications from coding and mathematical problem-solving to agentic functions and general conversational use cases. A key differentiator is its focus on user control and steerability, allowing users to define system prompts and alignment without external interference, addressing common issues with proprietary models like system prompt control, version changes, and data privacy.
Key Capabilities & Features
- General Purpose Reasoning: Trained for 3 epochs on 800k reasoning traces from the Dolphin-R1 dataset, aiming for comprehensive reasoning abilities.
- User Steerability: Provides full control over system prompts and alignment, enabling users to customize behavior and ethical guidelines.
- Versatile Applications: Designed for coding, math, agentic workflows, function calling, and general instruction-following.
- Data Privacy: Ensures user data remains private, as the model does not impose external ethics or collect queries.
- Optimized Temperature: Recommended low temperature settings (0.05 to 0.1) for optimal performance, avoiding issues like second-guessing.
Training & Data
The model leverages various open-source datasets, including those from OpenCoder-LLM, Microsoft (Orca), NousResearch (Hermes), AI-MO (NuminaMath), allenai (Tulu), and HuggingFaceTB (Smoltalk). It also incorporates CodeFeedback datasets from m-a-p. The development acknowledges contributions from Meta, Qwen, and OpenCoder, and utilizes RLHFlow for reward modeling and Deepseek-V3 for data augmentation.
Recommended Use Cases
- Local AI Development: Ideal for businesses and developers requiring a powerful, general-purpose model with full control over its behavior and data.
- Customizable AI Agents: Suitable for creating agents with specific tones, rules, and alignments via flexible system prompts.
- Coding & Math Assistance: Excels in generating code and solving mathematical problems.
- Function Calling: Capable of handling function calling tasks effectively.
- Privacy-Sensitive Applications: A strong choice for applications where data privacy and custom alignment are paramount.