Pinocchio-Lite-3B

Ministral-3-3B-Italian-KidsStories (finetuned)

Short description

A finetuned Italian instruction-following LLM intended to generate short, child-friendly educational stories in Italian. Built on top of mistralai/Ministral-3-3B-Instruct-2512 and finetuned primarily on an Italian children’s stories dataset (markod0925/Children-Stories-Collection-Italian) and additional curated, child-oriented prompts. Produces simple, age-appropriate narratives (3–5 paragraphs) that incorporate basic scientific concepts, memorable characters, and gentle moral lessons.

Model type: Causal language model (instruction-tuned, Italian)

Intended audience: Developers, educators, content creators, parents, and researchers wanting an Italian LLM for generating short educational stories for young children (with adult supervision).

Model overview

Optimized to:

• Produce simple Italian prose suitable for young children (vocabulary and sentence length aimed at early readers/listeners).
• Integrate elementary scientific concepts into narratives.
• Create memorable characters, dialogue-driven explanations, and a gentle twist with a life/science lesson.

Intended uses

Recommended:

• Generating short educational stories in Italian.
• Producing story prompts, outlines, or variations for human authors.
• Educational content drafts explaining simple scientific concepts.

Not recommended:

• Applications requiring guaranteed factual accuracy.
• Unsupervised distribution to children.
• High-stakes decision-making.

Limitations

• Hallucinations: May invent details.
• Bias & cultural sensitivity: Reflects training material biases.
• Safety for children: Not guaranteed to avoid inappropriate content.
• Evaluation gap: Further evaluation and filtering recommended for safety-critical use.

Training data & preprocessing

• Primary dataset: markod0925/Children-Stories-Collection-Italian (~17.3k examples).
• Preprocessing: Chat-style examples (prompt → user, text → assistant), tokenized with the model tokenizer.
• Filtering: Basic cleaning applied. Adult review recommended for PII or inappropriate content.

Safety considerations

• Human-in-the-loop: Adult review required before distribution to children.
• Automated filters: Use moderation pipeline for profanity, sexual/violent content.
• Prompt design: Avoid real-world medical, legal, or safety instructions.
• Disallowed uses: No impersonation of children or emotionally manipulative content.

How to use

Install

pip install transformers accelerate