Overview

This model, developed by walid-iguider, is an 8 billion parameter Llama-3 variant, fine-tuned from unsloth/llama-3-8b-bnb-4bit. It was trained using Unsloth for 2x faster processing and Huggingface's TRL library. The model is licensed under Apache-2.0.

Key Capabilities

• Italian Language Proficiency: Specifically fine-tuned and evaluated for performance in Italian.
• Efficient Training: Utilizes Unsloth for accelerated training, suggesting potential for efficient deployment.
• Benchmark Performance: Achieves an average normalized accuracy of 0.5334 across key Italian benchmarks, including:
  • hellaswag_it acc_norm: 0.6064
  • arc_it acc_norm: 0.4611
  • m_mmlu_it 5-shot acc: 0.5328

Good For

• Applications requiring a capable large language model for Italian text generation and understanding.
• Developers looking for an efficiently trained Llama-3 based model with a focus on Italian language tasks.
• Use cases where performance on Italian-specific benchmarks is critical.