Overview

e-palmisano/Phi3-ITA-mini-4K-instruct is a 4 billion parameter instruction-tuned language model developed by Enzo Palmisano. It is a specialized version of the microsoft/Phi-3-mini-4k-instruct model, specifically fine-tuned to excel in Italian language processing tasks. With a context length of 4096 tokens, this model is designed for efficient and accurate handling of Italian text.

Key Capabilities

• Italian Language Proficiency: Optimized for understanding and generating text in Italian.
• Instruction Following: Capable of following instructions for various natural language tasks.
• Performance on Italian Benchmarks: Achieves competitive scores on Italian-specific evaluation metrics, including hellaswag_it, arc_it, and m_mmlu_it.

Performance Metrics

The model's performance is evaluated on key Italian language benchmarks:

• hellaswag_it acc_norm: 0.6088
• arc_it acc_norm: 0.4440
• m_mmlu_it 5-shot acc: 0.5667
• Average Accuracy Normalized: 0.5398

For a comprehensive comparison, refer to the Leaderboard for Italian Language Models.

Good For

• Applications requiring a compact yet capable Italian language model.
• Tasks such as question answering, text generation, and summarization in Italian.
• Developers looking for a specialized model for Italian NLP projects.