Overview

e-palmisano/Qwen2-1.5B-ITA-Instruct is a 1.5 billion parameter Qwen2-based instruction-tuned language model developed by e-palmisano. It was fine-tuned from unsloth/Qwen2-1.5B-Instruct-bnb-4bit with a focus on enhancing its proficiency in the Italian language. The training involved continuous pretraining on 100k rows of the gsarti/clean_mc4_it dataset, followed by instruction-tuning on the FreedomIntelligence/alpaca-gpt4-italian dataset. This model benefits from being trained 2x faster using Unsloth and Huggingface's TRL library.

Key Capabilities

• Italian Language Proficiency: Specifically fine-tuned to improve understanding and generation in Italian.
• Instruction Following: Designed to respond effectively to instructions in Italian, leveraging the Alpaca-GPT4 Italian dataset.
• Efficient Training: Utilizes Unsloth for accelerated training, making it a resource-efficient option.

Performance

Evaluation on Italian language benchmarks shows the model's performance:

• hellaswag_it acc_norm: 48.05
• arc_it acc_norm: 32.68
• m_mmlu_it 5-shot acc: 46.89
• Average Accuracy Normalized: 42.57

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

Good for

• Applications requiring strong Italian language understanding and generation.
• Instruction-based tasks and chatbots in Italian.
• Developers looking for a specialized, efficiently trained Italian LLM.