normster/RealGuardrails-Llama3.1-8B-Instruct-SFT

Model Overview

normster/RealGuardrails-Llama3.1-8B-Instruct-SFT is an 8 billion parameter instruction-tuned model built upon the Llama 3.1 architecture, featuring a 32768 token context length. Developed by normster, this model's primary distinction lies in its specialized training on the RealGuardrails dataset. This dataset focuses on improving the model's ability to adhere to and prioritize system prompts, a critical aspect for robust AI applications.

Key Capabilities

• Enhanced System Prompt Adherence: The model is specifically trained to follow system-level instructions more reliably and consistently.
• Improved Precedence Handling: It demonstrates better understanding and application of instruction precedence within complex prompts.
• Instruction-Following: Optimized for general instruction-following tasks, with a particular emphasis on guardrail compliance.

Training Details

The model was fine-tuned via Supervised Fine-Tuning (SFT) on the systemmix split of the RealGuardrails dataset, comprising approximately 150,000 examples. Training was conducted using normster's custom torchllms library, ensuring a focused optimization for system prompt behavior. Key hyperparameters included an AdamW optimizer, a batch size of 128, a learning rate of 2e-5 with a cosine scheduler, and a maximum sequence length of 4096 during training.

Good For

• Applications requiring strict adherence to predefined rules and constraints.
• Scenarios where system prompts are used to enforce safety, style, or content guidelines.
• Developing AI agents that need to consistently follow complex, multi-layered instructions.