dphn/dolphin-2.9.3-llama-3-8b

TEXT GENERATIONConcurrency Cost:1Model Size:8BQuant:FP8Ctx Length:8kPublished:Jun 6, 2024Architecture:Transformer0.0K Cold

dphn/dolphin-2.9.3-llama-3-8b is an 8 billion parameter language model fine-tuned from Meta-Llama-3-8B, designed for conversational AI and agentic tasks. It leverages a diverse dataset including ShareGPT, SystemChat, Orca-Math, and ToolBench, optimizing it for complex instruction following, coding, and tool use. This model is particularly suited for applications requiring robust dialogue capabilities and problem-solving across various domains.

Loading preview...

dphn/dolphin-2.9.3-llama-3-8b: A Fine-Tuned Llama 3 for Conversational AI

This model is a fine-tuned version of Meta's Llama-3-8B, developed by dphn, focusing on enhancing its capabilities for conversational agents and complex instruction following. It was trained using the Axolotl framework, indicating a structured approach to fine-tuning.

Key Capabilities & Training

The model's training regimen involved a comprehensive and diverse set of datasets, including:

  • ShareGPT and SystemChat: For general conversational abilities and instruction adherence.
  • dolphin-coder-translate and dolphin-coder-codegen: Specifically targeting code generation and translation tasks.
  • m-a-p_Code-Feedback: Enhancing its understanding and application of code-related feedback.
  • Orca-Math: Improving mathematical reasoning and problem-solving.
  • ToolBench datasets (instruct, negative, react, tflan_cot): Crucial for developing tool-use capabilities and agentic reasoning, allowing the model to interact with external tools effectively.

Training was conducted over 3 epochs with a learning rate of 1e-05, utilizing a total batch size of 128 across 8 GPUs. The model's performance was monitored, showing a final validation loss of 0.5771.

Intended Use Cases

Given its training on a wide array of datasets, dphn/dolphin-2.9.3-llama-3-8b is well-suited for:

  • Advanced Chatbots: Capable of handling complex dialogues and following intricate instructions.
  • Code Generation & Translation: Assisting developers with programming tasks.
  • Agentic Applications: Where the model needs to reason, plan, and potentially use tools to achieve goals.
  • Mathematical Problem Solving: For tasks requiring numerical and logical reasoning.