Dolphin 3.0 Llama 3.1 8B Overview

Dolphin 3.0 is the latest iteration in the Dolphin series of instruct-tuned models, developed by Eric Hartford, Ben Gitter, BlouseJury, and Cognitive Computations. This 8 billion parameter model, based on the Llama 3.1 architecture, is engineered to be a versatile, general-purpose local model with a 32768 token context window.

Key Differentiators & Capabilities

User Control: Unlike proprietary models, Dolphin 3.0 emphasizes user control over system prompts and alignment, allowing developers to define ethics and guidelines without external interference.
General Purpose: Designed for a wide array of applications, including coding, mathematical problem-solving, agentic workflows, and function calling.
Data Privacy: Users maintain full control over their data, as queries are not visible to external providers.
Steerability: The model is highly steerable, enabling precise customization of tone, behavior, and rules via the system prompt.

Training & Data

The model leverages a diverse set of open-source datasets, including OpenCoder-LLM, Microsoft's Orca variants, NousResearch's function-calling data, AI-MO's mathematical datasets, and AllenAI's tulu-3-sft-mixture. Training was supported by various sponsors providing compute resources.

Performance Insights

While detailed evaluations are ongoing, initial Open LLM Leaderboard results show an average score of 24.97%, with notable performance in IFEval (76.21%) and BBH (27.63%).

Ideal Use Cases

Local Development: Businesses and developers requiring a powerful, general-purpose model for local deployment.
Custom Alignment: Projects where specific ethical guidelines or behavioral rules are critical and need to be user-defined.
Sensitive Data Handling: Applications involving proprietary or sensitive information where data privacy is paramount.
Coding & Math: Tasks requiring strong performance in code generation, understanding, and complex mathematical reasoning.