Model Overview

theprint/ReWiz-Llama-3.1-8B-v2 is an 8 billion parameter language model developed by theprint, fine-tuned from unsloth/meta-llama-3.1-8b-bnb-4bit. This model leverages the Llama 3.1 architecture and was trained with significant efficiency improvements using Unsloth and Huggingface's TRL library, resulting in 2x faster training times.

Key Characteristics

• Architecture: Llama 3.1-based, 8 billion parameters.
• Training Efficiency: Utilizes Unsloth for accelerated fine-tuning.
• Prompt Format: Optimized for an Alpaca-style prompt template to achieve best results.
• License: Released under the Apache-2.0 license.

Performance Benchmarks

Evaluated on the Open LLM Leaderboard, ReWiz-Llama-3.1-8B-v2 demonstrates a balanced performance across various metrics. Detailed results are available on the Open LLM Leaderboard and its specific evaluation page. Notable scores include:

• Avg.: 15.68
• IFEval (0-Shot): 23.73
• BBH (3-Shot): 23.77
• MMLU-PRO (5-shot): 25.67

Recommended Use Cases

This model is suitable for general-purpose language generation and understanding tasks where an 8B parameter model is appropriate. Its efficient training process suggests potential for further fine-tuning on specific downstream applications, particularly those benefiting from the Llama 3.1 base and Alpaca-style instruction following.