swap-uniba/LLaMAntino-3-ANITA-8B-Inst-DPO-ITA

LLaMAntino-3-ANITA-8B-Inst-DPO-ITA Overview

LLaMAntino-3-ANITA-8B-Inst-DPO-ITA is an 8 billion parameter instruction-tuned model based on Meta's Llama 3 architecture. Developed by Ph.D. Marco Polignano and the SWAP Research Group, this model is part of the ANITA project, which focuses on enhancing natural language interaction for the Italian language. It is designed to be a robust multilingual foundation, supporting both English and Italian.

Key Capabilities & Features

• Multilingual Support: Optimized for both English and Italian language use cases.
• Instruction-Tuned: Fine-tuned using QLoRA 4-bit on instruction-based datasets.
• DPO Alignment: Utilizes Direct Preference Optimization (DPO) with the mlabonne/orpo-dpo-mix-40k dataset to align with human preferences for helpfulness and safety.
• Context Length: Supports an 8K (8192 tokens) context window.
• Performance: Achieves an average score of 0.6160 on the Open Italian LLMs Leaderboard, with specific scores of 0.5714 on Arc_IT, 0.7093 on Hellaswag_IT, and 0.5672 on MMLU_IT.

Ideal Use Cases

• Italian NLP Research: Provides an improved model for researchers focusing on Italian language processing.
• Multilingual Applications: Suitable for applications requiring robust performance in both English and Italian.
• Further Fine-tuning: Serves as an excellent base model for domain-specific fine-tuning on Italian tasks.