TF1-EN-3M

홈페이지 · 논문 1편

# TF1-EN-3M: Three Million Synthetic Moral Fables for Open Language Models TF1-EN-3M is a large-scale synthetic dataset of 3,000,000 English-language moral fables, generated by instruction-tuned language models with no more than 8 billion parameters. The stories are aimed at child-friendly educational and moral reasoning applications and follow a consistent six-part narrative scaffold: character → trait → setting → conflict → resolution → moral. ## Dataset Characteristics - Size: 3 million stories (~1B tokens) - Format: JSON Lines, with detailed metadata including prompt elements, model configuration, generation time, token counts, and costs - Generation Models: Evaluated across 10 open-weight LLMs; final dataset generated using LLaMA-3.1-8B-Instruct for optimal quality and cost balance - Story Structure: Each story ends with an explicit moral and follows a template-driven structure - Target Audience: Designed primarily for children aged 4–7 (age group B), with simple vocabulary and accessible narratives ## Motivation Natural language processing lacks large, structured corpora of fables that combine creative storytelling with explicit moral lessons. Existing human-authored datasets like Aesop's Fables are limited in scale and diversity. TF1-EN-3M bridges this gap by: - Demonstrating that mid-sized open models can reliably generate coherent, instructive stories - Enabling research into value alignment, narrative intelligence, and low-resource model fine-tuning - Offering a reproducible, cost-efficient alternative to proprietary LLM pipelines ## Summary of Content Each entry in the dataset includes: - A structured prompt with narrative elements - The generated fable text - Metadata (model name, inference time, token usage, cost, etc.) - Quality assessments (via LLM-based scoring for grammar, creativity, moral clarity, and structure adherence) ## Use Cases TF1-EN-3M is suitable for a wide range of tasks and applications: - Training small or medium-sized LLMs for story generation or moral reasoning - Benchmarking models on tasks like moral inference, story-to-moral mapping, or story quality evaluation - Educational AI tools, such as interactive storytelling tutors or automated moral education platforms - Creative NLP research, including literary analysis and narrative generation - Multilingual extension by swapping out prompt elements for other languages ## Citation If you use TF1-EN-3M, please cite: ``bibtex @misc{nadas2025tf1en3m, title={TF1-EN-3M: Three Million Synthetic Moral Fables for Training Small, Open Language Models}, author={Mihai Nădaș and Laura Dioșan and Andreea Tomescu and Andrei Pișcoran}, year={2025}, eprint={2504.20605}, archivePrefix={arXiv}, primaryClass={cs.CL} } `` ## Dataset Access - Hugging Face: klusai/ds-tf1-en-3m - Generation & Evaluation Code: TinyFabulist GitHub