Model Overview

MaziyarPanahi/Llama-3-8B-Instruct-v0.9 is an instruction-tuned language model with 8 billion parameters, developed by MaziyarPanahi. It is an iteration built upon the Llama-3-8B-Instruct-v0.8 base model, focusing on enhanced instruction following capabilities.

Key Capabilities & Performance

This model demonstrates strong performance across various benchmarks, as evidenced by its evaluation on the Open LLM Leaderboard. As of March 6, 2024, it ranks as the 4th best-performing 8B model, achieving an average score of 73.29. Specific benchmark results include:

• AI2 Reasoning Challenge (25-Shot): 72.35
• HellaSwag (10-Shot): 88.17
• MMLU (5-Shot): 68.10
• TruthfulQA (0-shot): 64.67
• Winogrande (5-shot): 79.95
• GSM8k (5-shot): 66.49

Usage and Integration

The model utilizes the ChatML prompt template, ensuring compatibility with standard instruction-following formats. It can be readily integrated into applications using the Hugging Face transformers library, with support for torch.bfloat16 and device_map="auto" for efficient deployment. Quantized GGUF versions are also available for broader hardware compatibility.