💬 NLP

Children's English Reading Story Generation via Supervised Fine-Tuning of Compact LLMs with Controllable Difficulty and Safety

This paper demonstrates that compact 8B-parameter LLMs, when fine-tuned on expert-designed educational data, can generate children's English reading stories with controllable difficulty and safety that outperform zero-shot large models like GPT-4o and Llama 3.3 70B on difficulty metrics while remaining cost-effective for widespread educational use.

Original authors: Qian Shen (University of Florida, Gainesville, USA), Fanghua Cao (University of Florida, Gainesville, USA), Min Yao (University of Florida, Gainesville, USA), Shlok Gilda (University of Florida, Gaine

Published 2026-05-14

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Qian Shen (University of Florida, Gainesville, USA), Fanghua Cao (University of Florida, Gainesville, USA), Min Yao (University of Florida, Gainesville, USA), Shlok Gilda (University of Florida, Gainesville, USA), Bonnie J. Dorr (University of Florida, Gainesville, USA), Walter L. Leite (University of Florida, Gainesville, USA)

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 are a teacher trying to give every child in your class a custom-made storybook. You want the story to be about a specific topic (like "cats" or "rainbows"), but you also need the words to be just right: not too hard for a kindergartener, but not so simple that a second-grader gets bored. Plus, the story must be safe and free of anything scary or mean.

Doing this by hand for hundreds of kids is impossible. You might think, "Let's just ask a super-smart AI robot to write them!" But here's the problem: the biggest, smartest AI robots (like the ones running on massive supercomputers) are too expensive to run in every classroom, and when you ask them to follow strict rules, they often get confused. They might write a story that is too hard, or they might accidentally include a scary monster when you asked for a friendly one.

The Big Idea: The "Compact" Robot
This paper is about training smaller, cheaper, and more manageable AI robots (called "8B models") to be the perfect storybook writers. The researchers wanted to see if they could teach these smaller robots to follow strict rules better than the giant, expensive ones.

Think of the giant AI as a famous, overworked chef who is great at cooking complex gourmet meals but struggles to follow a simple recipe for a child's lunch. The smaller AI is like a local baker who needs specific training to learn exactly how to make the perfect child-friendly sandwich.

How They Did It (The Training Camp)
The researchers took an existing, expert-designed reading curriculum (a list of 129 lessons with specific sounds and words to teach). They first asked the giant AI to write 2,580 stories based on these lessons. Even though the giant AI wasn't perfect, it provided a good starting point.

Then, they put three different "baker" robots through a special training camp using three different methods:

The "Just Do It" Method (Baseline): They simply showed the robots the lesson and the story, saying, "Copy this style."
The "Best of the Best" Method (Good Stories): They filtered the stories and only showed the robots the highest-quality ones, thinking, "If you learn from the best, you'll be the best."
The "Reward System" Method (Rewarded SFT): This was the star of the show. They gave the robots a score for every story based on how easy it was to read, how safe it was, and how logical it sounded. If the robot wrote a good story, it got a "gold star" (a reward). If it wrote a bad one, it got a lower score. The robot learned to chase the gold stars.
The "Mistake Simulator" Method: They used a super-smart AI to pretend to be a child making reading mistakes, then taught the robots how to handle those errors.

The Results: The Small Robot Wins
When they tested the results, the "Reward System" method was the clear winner. Here is what they found:

Difficulty Control: The stories written by the trained small robots were actually easier and more appropriate for young children (K-2) than the stories written by the giant, zero-shot AI. The giant AI kept using big, fancy words, while the small robots learned to stick to simple words.
Safety: The small robots were very safe. Almost all the stories were free of mean language or scary content.
Creativity vs. Rules: Usually, small robots struggle to be creative while following strict rules. But with the right training (the Reward System), these small robots managed to write stories that were both simple enough for kids and still made sense as stories.

The "Bad" Story vs. The "Good" Story
The researchers looked closely at the stories.

The "Good" Story: It flowed like a gentle stream. The events happened in order, the words were simple, and the characters stayed consistent. It was easy for a child to follow.
The "Bad" Story: It was like a rollercoaster that went off the tracks. It jumped between unrelated ideas, used big words like "insatiable" (which a 6-year-old wouldn't know), and sometimes included weird, slightly mean comments like "you're too fat."

The Catch (Limitations)
While the small robots did a great job, they aren't perfect yet.

Repetition: The robots sometimes got a little stuck in a rut, using the same names (like "Sam" and "Pam") or places over and over again. This happened because the training data they learned from had those same patterns.
Size: Even though these are "small" robots, they are still too big to run on a regular tablet or phone. They need a decent computer to work.
Human Check: The robots still need a human (a teacher or parent) to look over the stories before giving them to a child, just to make sure nothing slipped through the cracks.

The Bottom Line
This paper proves that you don't need a billion-dollar supercomputer to generate great, safe, and simple reading stories for children. By using a smart training method (the "Reward System"), you can teach a smaller, cheaper AI to do the job better than the giant, expensive one. This means teachers and parents could eventually run these story generators on their own computers at home or in the classroom, creating endless, custom stories that are perfectly tuned to a child's reading level.

Technical Summary: Children's English Reading Story Generation via Supervised Fine-Tuning of Compact LLMs

Problem Statement

While Large Language Models (LLMs) offer promise for generating personalized educational content, their application in children's reading instruction faces two critical barriers:

Pedagogical Misalignment: State-of-the-art models (e.g., GPT-4o, Llama 3.3 70B) struggle to strictly adhere to early-elementary pedagogical constraints. Specifically, they often fail to restrict lexical and syntactic difficulty to target K–2 reading levels without compromising narrative coherence.
Accessibility and Cost: The operational costs of closed-source APIs and the massive hardware requirements for deploying 70B+ parameter models locally hinder widespread adoption in classrooms and homes.

The core challenge is adapting compact, local LLMs (under 10B parameters) to generate stories that balance narrative creativity with strict rule-following (vocabulary boundaries, syntax simplification, and safety), a task where small models typically suffer from mode collapse or logical fragmentation under multi-dimensional constraints.

Methodology

The authors conducted a systematic empirical comparison of various fine-tuning strategies using three 8B-parameter LLMs: Llama 3 8B, Apertus 8B-instruct-2509, and Granite 3.3 8b-instruct.

Data Construction

Source: The study utilized a K–2 English reading curriculum from the University of Florida Literacy Institute (UFLI), consisting of 129 lessons with specific target phonemes.
Training Data: Using zero-shot prompting, GPT-4o and Llama 3.3 70B generated 2,580 stories (10 per lesson). Despite their limitations in strict constraint satisfaction, these stories provided the necessary linguistic quality and narrative coherence to serve as training targets.
Preprocessing: The 2,580 stories were processed into four distinct experimental datasets:
1. Baseline: All 2,580 stories used directly.
2. "Good Stories": A filtered subset of 996 stories that performed above the corpus mean on five metrics (Spache readability, perplexity, coherence, syntactic complexity, toxicity).
3. Rewarded SFT: All 2,580 stories used, but with a scalar reward assigned to each example based on a normalized average of the five metrics. The cross-entropy loss was reweighted by these rewards during training.
4. SFT with Simulated Errors: Inspired by Self-Instruct, the authors used GPT-OSS-120B to simulate 3–8 phonemic mispronunciations typical of early readers for each story. These simulated errors were concatenated with lesson phonemes as input to the model.

Fine-Tuning Configuration

Technique: Parameter-Efficient Fine-Tuning (PEFT) using QLoRA (4-bit NF4 quantization + LoRA adapters).
Hyperparameters: LoRA rank=32, $\alpha$ =64, dropout=0.1, learning rate=1e-4, batch size=4, 5 epochs, warmup ratio=0.03, optimizer=AdamW.
Inference: Nucleus sampling (top-p=0.9) with temperature 0.8.

Evaluation Metrics

Stories were evaluated on difficulty, safety, and diversity using:

Spache Readability Score: Targeting K–2 levels (score < 3.0).
LM Perplexity (LM-PPL): Measuring fluency and naturalness (using GPT-2).
Coherence: Average shared named entities between consecutive sentences.
Syntactic Complexity: Composite metric of maximum dependency distance and subordinate clauses.
Toxicity: Measured via Detoxify (targeting near-zero scores).
Self-BLEU: Measuring inter-story repetition and generative diversity.

Key Results

Quantitative Findings

Difficulty Control: Fine-tuned 8B models, particularly those trained with Rewarded SFT, achieved significantly lower Spache readability scores (indicating easier text) compared to zero-shot GPT-4o and Llama 3.3 70B. The best-performing models (Granite 3.3 and Apertus) reached average Spache scores around 2.3–2.7, approaching the second-grade level.
Safety: All generated stories exhibited toxicity scores very close to 0, indicating minimal safety risks.
Stability: The Rewarded SFT approach produced highly stable outputs, with almost no hallucinations, garbled text, or story-irrelevant content. Other experiments showed less than 10% undesirable outputs.
Statistical Significance: Welch's T-tests confirmed that the improvements in coherence, syntactic complexity, Spache readability, and LM-PPL between the fine-tuned 8B models (specifically Rewarded SFT) and the zero-shot 70B models were statistically significant ( $p < 0.001$ , Cohen's $d > 0.8$ ).
Model Performance: Among the three 8B models, Llama 3 8B consistently underperformed compared to Granite 3.3 and Apertus.

Qualitative Findings

Outliers: Some stories in the "Simulated Errors" experiment showed higher perplexity due to brevity or appended irrelevant content.
Toxicity Cases: A small number of stories (<5 per dataset) contained potentially harmful language (e.g., body shaming, unfriendly characters), highlighting the need for human oversight.
Repetition: The models exhibited a tendency to reuse specific names (e.g., "Sam," "Pam"), animals, and locations across different lessons. The authors attribute this to the generative preferences encoded in the source training data (the GPT-4o/Llama 3.3 70B stories), suggesting that fine-tuning transfers these specific patterns to the compact models.

Key Contributions

Empirical Validation of Compact Models: The study demonstrates that with appropriate fine-tuning pipelines (specifically Rewarded SFT), compact 8B models can generate children's stories with controllable difficulty and safety that match or surpass zero-shot 70B models.
Rewarded SFT Strategy: The introduction of a reward-weighted loss function based on multiple pedagogical metrics proved effective in stabilizing generation and improving alignment with K–2 constraints without requiring full RLHF pipelines.
Cost-Effective Accessibility: The work provides a pathway for democratizing personalized educational content by enabling deployment on local hardware (under 10B parameters) rather than relying on expensive APIs or massive local clusters.

Significance and Claims

The paper claims that its approach offers a viable, equitable solution for AI-driven reading interventions. By prioritizing controllability over scale, the proposed method allows educators to target specific reading levels and error patterns using affordable, compact models.

The authors emphasize that while the fine-tuned 8B models show significant promise for K–2 story generation, they are not yet fully autonomous. The study advocates for these models to be used as assistive tools under human supervision (teachers/parents) to ensure safety and pedagogical appropriateness. The work highlights that future iterations require more diverse data, including authentic child reading errors and expert human ratings, to further refine the models and address limitations such as content repetition and the need for generalization to unseen phoneme constraints.