Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

TF1-EN-3M: Three Million Synthetic Moral Fables for Training Small, Open Language Models

Mihai Nadas, Laura Diosan, Andrei Piscoran, Andreea Tomescu

TL;DR

TF1-EN-3M tackles the lack of open, large-scale moral fables by generating 3 million stories with instruction-tuned models (≤8B) using a six-slot template to ensure coherent, morals-driven narratives. The authors present a combinatorial prompt engine and a hybrid evaluation framework combining an LLM critic and reference-free metrics to select suitable open-weight models, with Llama-3.1-8B-Instruct identified as the best trade-off. They release the dataset, generation code, evaluation scripts, and metadata, enabling reproducibility and cost benchmarking, and demonstrate practical utility for instruction following, narrative intelligence, and value-aligned educational AI on commodity hardware. Overall, TF1-EN-3M provides a scalable, accessible resource for research into moral storytelling, safe AI education, and downstream fine-tuning of smaller models for on-device generation and moral reasoning benchmarks.

Abstract

Moral stories are a time-tested vehicle for transmitting values, yet modern NLP lacks a large, structured corpus that couples coherent narratives with explicit ethical lessons. We close this gap with TF1-EN-3M, the first open dataset of three million English-language fables generated exclusively by instruction-tuned models no larger than 8B parameters. Each story follows a six-slot scaffold (character -> trait -> setting -> conflict -> resolution -> moral), produced through a combinatorial prompt engine that guarantees genre fidelity while covering a broad thematic space. A hybrid evaluation pipeline blends (i) a GPT-based critic that scores grammar, creativity, moral clarity, and template adherence with (ii) reference-free diversity and readability metrics. Among ten open-weight candidates, an 8B-parameter Llama-3 variant delivers the best quality-speed trade-off, producing high-scoring fables on a single consumer GPU (<24 GB VRAM) at approximately 13.5 cents per 1,000 fables. We release the dataset, generation code, evaluation scripts, and full metadata under a permissive license, enabling exact reproducibility and cost benchmarking. TF1-EN-3M opens avenues for research in instruction following, narrative intelligence, value alignment, and child-friendly educational AI, demonstrating that large-scale moral storytelling no longer requires proprietary giant models.

TF1-EN-3M: Three Million Synthetic Moral Fables for Training Small, Open Language Models

TL;DR

TF1-EN-3M tackles the lack of open, large-scale moral fables by generating 3 million stories with instruction-tuned models (≤8B) using a six-slot template to ensure coherent, morals-driven narratives. The authors present a combinatorial prompt engine and a hybrid evaluation framework combining an LLM critic and reference-free metrics to select suitable open-weight models, with Llama-3.1-8B-Instruct identified as the best trade-off. They release the dataset, generation code, evaluation scripts, and metadata, enabling reproducibility and cost benchmarking, and demonstrate practical utility for instruction following, narrative intelligence, and value-aligned educational AI on commodity hardware. Overall, TF1-EN-3M provides a scalable, accessible resource for research into moral storytelling, safe AI education, and downstream fine-tuning of smaller models for on-device generation and moral reasoning benchmarks.

Abstract

Moral stories are a time-tested vehicle for transmitting values, yet modern NLP lacks a large, structured corpus that couples coherent narratives with explicit ethical lessons. We close this gap with TF1-EN-3M, the first open dataset of three million English-language fables generated exclusively by instruction-tuned models no larger than 8B parameters. Each story follows a six-slot scaffold (character -> trait -> setting -> conflict -> resolution -> moral), produced through a combinatorial prompt engine that guarantees genre fidelity while covering a broad thematic space. A hybrid evaluation pipeline blends (i) a GPT-based critic that scores grammar, creativity, moral clarity, and template adherence with (ii) reference-free diversity and readability metrics. Among ten open-weight candidates, an 8B-parameter Llama-3 variant delivers the best quality-speed trade-off, producing high-scoring fables on a single consumer GPU (<24 GB VRAM) at approximately 13.5 cents per 1,000 fables. We release the dataset, generation code, evaluation scripts, and full metadata under a permissive license, enabling exact reproducibility and cost benchmarking. TF1-EN-3M opens avenues for research in instruction following, narrative intelligence, value alignment, and child-friendly educational AI, demonstrating that large-scale moral storytelling no longer requires proprietary giant models.
Paper Structure (42 sections, 6 equations, 1 figure, 6 tables)

This paper contains 42 sections, 6 equations, 1 figure, 6 tables.

Table of Contents

  1. Introduction
  2. Research Questions
  3. Contributions
  4. Prompt design and dataset generation
  5. Prompt Schema and Value Space
  6. Prompt Construction:
  7. System Message Design:
  8. Value Space and Prompt Design Variables
  9. Full System Message and Prompt Template
  10. Evaluation Methodology
  11. LLM-based Critic
  12. Age Group Classification
  13. Reference-Free Metrics
  14. LLM Evaluation and Comparison with Related Work
  15. Experimental Setup
  16. ...and 27 more sections

Figures (1)

  • Figure 1: Full pipeline for generating TF1-EN-3M.