SerFabio89/italian-sft-dataset

Italian High-Quality SFT Dataset

This dataset is a diverse, high-quality, fully Italian instruction-tuning dataset designed for fine-tuning Large Language Models (LLMs). It provides a comprehensive set of instructions to enhance model helpfulness, logical reasoning, and instruction-following capabilities in Italian.

Dataset Details

• Language: Italian
• Format: Multi-turn and single-turn instructions, structured data, logical reasoning, programming, and long-context QA.
• Generation Method: The dataset was synthetically generated and distilled using a carefully weighted mix of state-of-the-art open-source models (including a mix of Llama, Qwen, and other open-source models).

Dataset Structure

The dataset follows a standard conversation format, making it directly compatible with Hugging Face's apply_chat_template or standard ChatML formatting.

• id: Unique identifier for the generation
• category: The task category (e.g. scrittura_professionale, spiegazioni_didattiche, ragionamento_logico_matematico, programmazione_python)
• length_band: Target length band of the generation (short, medium, long)
• messages: List of messages, where each message contains role (user or assistant) and content.
• thinking: Boolean flag indicating if the assistant output contains a reasoning trace.

Generation Pipeline

• The prompts were mined using a combination of Magpie generation, self-instruct meta-prompting, and curated seed files.
• The responses were generated by an ensemble of high-quality open-source teachers.
• The generated dataset underwent extensive offline quality filtering including semantic embedding deduplication, BGE reranker scoring, and language identification to ensure high-quality Italian outputs.
• All tokens were validated to drop records containing excessive unknown or malformed tokens.