Aether Drake S F T

Introduction

AetherDrake-SFT is an advanced AI model optimized for logical reasoning, multi-step problem-solving, and decision-making tasks. It uses a structured system prompt and iterative thought processes to deliver high-quality responses.

Architecture

The model operates with a series of structured steps:

• Initial Thought: Uses <Thinking> tags for step-by-step reasoning.
• Self-Critique: Evaluates initial responses using <Critique> tags, focusing on accuracy, clarity, completeness, and areas for improvement.
• Revision: Refines responses as needed within <Revising> tags.
• Final Response: Presents a polished response within <Final> tags.
• Tag Innovation: Introduces new tags for enhanced clarity and structure.

Training

AetherDrake-SFT is trained using:

• Frameworks: Unsloth, Hugging Face Transformers, and the TRL library.
• Datasets: Focused on reasoning tasks, including logical puzzles and commonsense reasoning scenarios.
• Hardware Efficiency: Utilizes bnb-4bit precision for reduced memory usage and faster development cycles.

Guide: Running Locally

1. Installation: Clone the repository and install the required dependencies using pip install -r requirements.txt.
2. Setup: Configure your environment by setting up API keys if needed.
3. Run the Model: Execute the script to start generating text.
4. Cloud GPUs: For optimal performance, consider using cloud GPU services like AWS, Google Cloud, or Azure.

License

AetherDrake-SFT is distributed under the Apache 2.0 license, allowing for use, modification, and sharing per the license terms.