vicgalle/ConfigurableSOLAR-10.7B

Overview

vicgalle/ConfigurableSOLAR-10.7B is a 10.7 billion parameter language model that implements a novel approach called configurable safety tuning (CST). Developed by vicgalle, this model allows users to define its behavior, including safety and helpfulness levels, by providing specific system prompts. This flexibility is achieved through fine-tuning on the vicgalle/configurable-system-prompt-multitask dataset.

Key Capabilities

• Dynamic Behavior Configuration: Users can switch between different operational modes (e.g., helpful and harmless, completely uncensored, unbiased and truthful) using tailored system prompts.
• Persona Role-playing: The model can adopt specific role-played personas based on the system prompt provided.
• Research Focus: This model is a practical application of the research presented in the paper "Configurable Safety Tuning of Language Models with Synthetic Preference Data" (arXiv:2404.00495).

Use Cases

• Research and Experimentation: Ideal for exploring the impact of safety tuning and system prompt engineering on LLM behavior.
• Applications Requiring Adaptable AI: Suitable for scenarios where the AI's persona or safety guardrails need to be dynamically adjusted by the user or application.

Performance Notes

According to the Open LLM Leaderboard, the model achieved an average score of 19.05. Specific scores include 51.00 on IFEval (0-Shot) and 27.45 on BBH (3-Shot). Users should review the detailed leaderboard results for specific task performance.