Dolphin 3.0 Qwen 2.5 3B Overview

Dolphin 3.0 Qwen 2.5 3B is a 3.1 billion parameter instruction-tuned model, part of the Dolphin 3.0 series, developed by Eric Hartford, Ben Gitter, BlouseJury, and Cognitive Computations. It is engineered to be a versatile local model, capable of handling a wide range of tasks including coding, mathematical problem-solving, agentic workflows, function calling, and general conversational interactions.

Key Differentiators & Capabilities

• User Control and Steerability: Unlike many proprietary models, Dolphin 3.0 emphasizes user control. It allows users to define the system prompt and alignment, ensuring that the model's behavior and ethical guidelines are dictated by the user's specific application needs. This provides full control over data privacy and model responses.
• General Purpose Design: Aims to serve as a comprehensive tool for various applications, similar to larger commercial models but with the advantage of local deployment and user customization.
• Chat Template: Utilizes the ChatML format for its chat template, providing a structured way to interact with the model.

Training & Data

The model's training leveraged a diverse set of open-source datasets, including contributions from OpenCoder-LLM, Microsoft (orca-agentinstruct, orca-math-word-problems), NousResearch (hermes-function-calling), AI-MO (NuminaMath-CoT, NuminaMath-TIR), allenai (tulu-3-sft-mixture), and HuggingFaceTB (smoltalk). The development also benefited from the use of RLHFlow for dataset filtering and Deepseek-V3 for data augmentation.