Dolphin 2.9 Llama 3 8b Overview

Dolphin 2.9 Llama 3 8b is a fine-tuned language model developed by Eric Hartford, Lucas Atkins, Fernando Fernandes, and Cognitive Computations, built upon Meta's Llama-3-8B architecture. It was trained using a full-weight fine-tuning approach with a 4k sequence length, taking 2.5 days on 8x L40S GPUs. The model utilizes the ChatML prompt template format and was trained on a diverse dataset including ShareGPT, Ultrachat, and various coding and agentic datasets.

Key Capabilities

Instruction Following: Excels at understanding and executing complex instructions.
Conversational AI: Designed for natural and engaging dialogue.
Coding Skills: Possesses capabilities for code generation and translation.
Agentic Abilities: Includes initial support for agentic workflows and function calling.
Uncensored Nature: The model is uncensored, offering high compliance with user requests, and requires users to implement their own alignment layers for ethical use.

Use Cases

General-Purpose AI Assistant: Suitable for a wide range of tasks due to its broad capabilities.
Developer Tools: Can be integrated into applications requiring code generation or function calling.
Research and Experimentation: Its uncensored nature makes it valuable for exploring AI behavior without inherent biases, with the caveat that users are responsible for content generated.

Note: A known bug in the SystemConversations dataset may cause the model to overly discuss the "SYSTEM MESSAGE." It is recommended to include a directive in the system message to mitigate this behavior.