💬 NLP

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

The paper introduces TF1-EN-3M, a novel open dataset of three million synthetic English moral fables generated by small, open-weight language models using a structured six-slot scaffold and a reproducible evaluation pipeline, demonstrating that high-quality, large-scale moral storytelling can be achieved without relying on proprietary giant models.

Original authors: Mihai Nadas, Laura Diosan, Andrei Piscoran, Andreea Tomescu

Published 2026-05-06

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Mihai Nadas, Laura Diosan, Andrei Piscoran, Andreea Tomescu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you want to teach a child the difference between right and wrong. For centuries, people have used fables—short stories about talking animals or simple characters that end with a clear lesson (like "Honesty is the best policy"). These stories are like moral GPS systems: they guide the listener from a confusing situation to a clear ethical destination.

However, in the world of computer science (specifically Natural Language Processing), researchers hit a wall. They needed a massive library of these stories to teach computers how to understand and tell moral tales, but existing collections (like Aesop's Fables) were too small. They also didn't want to spend millions of dollars using giant, expensive AI models to create them.

Enter TF1-EN-3M. Think of this project as a massive, automated story factory built by researchers in Romania. Here is how they did it, broken down simply:

1. The Recipe Book (The Prompt Engine)

Instead of asking a computer to "write a story" and hoping for the best, the researchers built a strict recipe. They created a "scaffold" with six specific slots, like a fill-in-the-blank worksheet:

Character: (e.g., a Fox)
Trait: (e.g., Greedy)
Setting: (e.g., A busy farm)
Conflict: (e.g., Steals food)
Resolution: (e.g., Gets caught)
Moral: (e.g., "Greed leads to trouble")

They didn't just write one recipe; they created a combinatorial engine. Imagine having 100 different characters, 100 different traits, and 100 different settings. If you mix and match them, you get millions of unique story ideas. They used this engine to generate 3 million unique prompts.

2. The Bakers (The AI Models)

The researchers didn't use the most expensive, super-sized AI models (the "giant chefs" that cost a fortune to run). Instead, they tested ten different "compact" AI models (ranging from 1 billion to 8 billion parameters). Think of these as home bakers rather than industrial factories.

They asked these home bakers to follow the strict recipes. To see which baker was the best, they set up a panel of judges.

The Judges: Instead of hiring expensive human critics, they used three other AI models to grade the stories on grammar, creativity, how clear the moral was, and whether the baker followed the recipe.
The Winner: They found that a specific model called Llama-3.1-8B was the "Goldilocks" choice. It wasn't the absolute highest scorer in every single category, but it was the best at balancing quality with cost. It could write high-quality stories on a regular computer (consumer hardware) for about $0.135 per 1,000 stories.

3. The Result: A Library of 3 Million Stories

The result is the TF1-EN-3M dataset.

Size: 3,000,000 English fables.
Content: Every story is short (about the length of a bedtime story), uses simple words suitable for children aged 4–7, and ends with a clear moral lesson.
Safety: The researchers checked the stories and found they are safe. While they contain mild conflicts (like stealing or fighting), these are always used to teach a lesson, not to be harmful.
Cost: The entire library was created for a total cost of $405.

4. Why This Matters (According to the Paper)

The paper claims this is a breakthrough because it proves you don't need a supercomputer to create massive, high-quality educational content.

Reproducibility: They released the code, the data, and the "recipes" for free. Anyone can run the factory again and get the exact same results.
Efficiency: It shows that small, open-source models can do the job of "moral storytelling" just as well as the giant, expensive ones.
Use Cases: The paper suggests this dataset can be used to train smaller AI models to become better at telling stories, understanding morals, or helping with educational tools for children.

The Catch (Limitations)

The authors are honest about the limits. Because the stories are built from a strict recipe, they might feel a bit repetitive (like a song with the same chord progression). They also noted that the stories are based on Western fable traditions (like Aesop), so they might not reflect every culture's view on morality.

In a nutshell: The researchers built a machine that mixes and matches story ingredients to bake 3 million moral fables. They proved you can do this cheaply and quickly using small, open AI models, and they handed the recipe and the cookies to the world for free.

Technical Summary: TF1-EN-3M

Problem Statement

While moral stories (fables) are a proven medium for transmitting values, modern Natural Language Processing (NLP) lacks a large-scale, structured corpus that couples coherent narratives with explicit ethical lessons. Existing classical collections (e.g., Aesop's Fables) are too small for data-driven modeling, and current synthetic data generation often lacks the specific structural rigor required for moral reasoning tasks. Furthermore, there is a need to determine whether compact, open-weight language models (LLMs) running on consumer hardware can reliably generate high-quality, instructive narratives without relying on proprietary giant models or expensive API-based evaluation infrastructure.

Methodology

1. Prompt Design and Data Generation

The authors developed a combinatorial prompt engine to generate 3,000,000 unique English fables. The generation process relies on a six-slot scaffold derived from narrative theory:

Character: Protagonist (e.g., animal or human archetype).
Trait: A defining attribute (e.g., greedy, kind).
Setting: Contextual location.
Conflict: The central challenge or dilemma.
Resolution: The outcome demonstrating a moral judgment.
Moral: The explicit ethical lesson.

The input space is defined by the Cartesian product of curated lists for each slot ( $T = n \times m \times k \times c \times r \times l$ ), where each parameter was set to 100 options. To ensure quality, the system employs:

System Messages: Framing the task as moral fable writing with specific constraints on tone, audience (targeting ages 4–7), and structure.
Stylistic Constraints: Instructions to "show, don't tell," avoid character names, use simple vocabulary, and maintain a concise length (~250 words).
Sampling Safeguards: Uniform random sampling with deduplication, frequency filtering for common pairings, and coverage balancing to ensure each slot appears equally.

2. Evaluation Framework

To select the optimal generator, the authors evaluated ten open-weight, instruction-tuned LLMs (ranging from 1B to 32B parameters) using a multi-faceted pipeline:

Multi-Judge LLM Panel: Three open-weight models from distinct families (Granite 4.1 30B, EXAONE 3.5 32B, and Granite 3.3 8B) scored 1,000 generated fables each on four dimensions: Grammar & Style, Creativity, Moral Clarity, and Prompt Adherence.
Reference-Free Metrics: Corpus-level statistics including Self-BLEU (diversity), Distinct-1 (lexical richness), and Flesch Reading Ease (readability).
Age Group Classification: An LLM classifier assigned each story to one of five age brackets (A–E) to verify alignment with the target demographic (4–7 years).
Composite Scoring: A weighted score ( $W_m$ ) was calculated, prioritizing Prompt Adherence (0.35) and Moral Clarity (0.20), followed by Grammar, Creativity, and the reference-free metrics.

3. Model Selection

Ten candidates were tested, including variants of Llama-3, Mistral, Qwen, Falcon, and Phi. While Llama-3.1-Tulu-3-8B achieved the highest composite score (0.957) based on raw metrics, Llama-3.1-8B-Instruct was selected for the full-scale generation. This decision was driven by its superior alignment with the target age group (92% of its outputs were classified as suitable for ages 4–7) and its favorable quality-cost trade-off on consumer hardware.

Key Contributions

TF1-EN-3M Dataset: The release of the first large-scale open dataset of synthetic fables in English, containing 3 million stories. Each entry includes the narrative, the structured prompt, and generation metadata (token counts, latency, cost, hardware specs).
Efficient Open-Weight Generation: Demonstration that high-quality, instructive content can be generated at scale using models with 1–8 billion parameters. The entire corpus was generated for approximately $405.76 (approx. $0.135 per 1,000 fables) on consumer-grade hardware.
Reproducible Multi-Judge Evaluation: A fully reproducible framework using a panel of open-weight LLM judges from distinct families, combined with reference-free metrics. This approach avoids reliance on proprietary evaluation infrastructure while mitigating self-preference bias.

Results

Dataset Quality: The final corpus exhibits high lexical diversity (Distinct-1 = 0.452) and readability (Flesch Reading Ease score of 78.9, corresponding to a 5.5-grade level), aligning with the target audience of young readers. Near-duplicate detection confirmed the combinatorial approach produced genuinely distinct narratives.
Model Performance: Modern 7–8B instruction-tuned models converged in quality, with top performers (Llama-3.1-Tulu-3-8B, Falcon3-7B-Instruct, Llama-3.1-8B-Instruct) scoring highly across all axes. Smaller models (e.g., 1B–3B) showed strong potential for low-latency applications.
Evaluation Reliability: The open-weight judge panel showed strong rank-order agreement with a proprietary judge (GPT-o4-mini), with a Kendall's $\tau$ of 0.78, validating that open models can substitute for proprietary evaluation infrastructure in model selection.
Content Safety: A safety scan of 5,000 samples revealed that 87.8% contained no flagged keywords. The remaining flagged terms (e.g., "steal," "fight") appeared in didactic contexts reinforcing prosocial values, with no severe content detected.

Significance and Claims

The paper claims that TF1-EN-3M opens new avenues for research in instruction following, narrative intelligence, value alignment, and child-friendly educational AI. The authors assert that large-scale moral storytelling does not require proprietary giant models or expensive evaluation infrastructure.

By providing a resource that couples coherent narratives with explicit ethical lessons, the dataset enables:

Fine-tuning of smaller models for fable generation and moral reasoning evaluation.
Training of classifiers or QA models on moral content.
Benchmarking of generative models on narrative alignment and commonsense reasoning.

The work emphasizes that with focused prompt engineering and curated pipelines, mid-sized open models can produce diverse, ethically themed narratives, making value-laden AI more accessible and socially grounded. The authors explicitly note limitations, acknowledging that the dataset reflects Western fable traditions and may lack the nuance for complex contemporary ethical dilemmas, suggesting future work should explore multi-layered morals and diverse cultural contexts.