Zefiro-7b-sft-qlora-ITA-v0.5: Italian Language Fine-Tuned Model

Zefiro is a 7 billion parameter GPT-like model developed by giux78, fine-tuned for the Italian language. It is based on the mistralai/Mistral-7B-v0.1 architecture and draws inspiration from the Zephyr model's approach, adapted for Italian. The project's goal is to create open-source models and datasets tailored for the Italian language, with Zefiro being the initial experiment.

Key Capabilities

• Italian Language Focus: Primarily designed and optimized for generating text in Italian.
• Fine-tuned for Conversations: Utilizes a Supervised Fine-Tuning (SFT) approach, making it suitable for conversational tasks.
• Synthetic Data Training: Trained on a filtered and preprocessed version of the UltraChat-ITA dataset, which consists of synthetic dialogues generated by ChatGPT.
• Base Model for Specific Tasks: Intended to be used as a base model for more specialized Italian conversational applications.

Intended Uses and Limitations

Zefiro is well-suited for general Italian language generation and conversational AI. However, it has not undergone human preference alignment for safety (like RLHF) or in-the-loop filtering, meaning it may produce problematic outputs if prompted. The model's training data, while focused on Italian, is derived from synthetic sources, and the base Mistral model's original training corpus composition is not fully detailed. Users should be aware of these limitations, particularly regarding safety and potential biases inherent in the synthetic training data and the base model.