SteelStorage/llama-3-cat-8b-instruct-v1

Overview

SteelStorage/llama-3-cat-8b-instruct-v1 is an 8 billion parameter Llama 3 instruction-tuned model, developed by SteelSkull with dataset preparation by Dr. Kal'tsit. This model is specifically fine-tuned to prioritize system prompt fidelity, helpfulness, and character engagement, aiming for deep immersion in role-play scenarios.

Key Capabilities

• System Instruction Fidelity: Designed to adhere strictly to system prompts.
• Chain of Thought (COT): Capable of generating detailed, step-by-step reasoning, though this behavior is primarily driven by system card instructions rather than inherent fine-tuning.
• Character Immersion: Optimized for maximum character engagement and role-play.
• Helpfulness: Provides helpful information, with a particular focus on biosciences and general science, drawing from health-related data for detailed diagnoses.

Training Details

The model was trained on a filtered Hugging Face dataset of instruction-response pairs, with a GPT model used to establish a standard for high-quality responses. The dataset was further refined for length and COT responses, and health-related data from Chat Doctor was included, favoring detailed and step-by-step diagnoses. Training involved 4 epochs over 6 days on a single A100 GPU.

Performance

Evaluations on the Open LLM Leaderboard show an average score of 64.74, with notable scores in HellaSwag (79.20) and Winogrande (75.93).