Toxicity of stories generated by the fine-tuned 8B LLMs is very close to zero, with only a minimal number of stories posing potential safety risk¶

Claim · All claims
Evidence · 1 study · design r2 · q2
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person · recheck by 2029: tool and model accuracy changes quickly

Subclaims¶

q2 i? Higher toxicity outlier stories were traced to specific content such as insults, disgust-related language, and unfriendly characters. → Qian Shen 2026

Evidence¶

Qian Shen 2026¶

Qian Shen, Fanghua Cao, Min Yao, Shlok Gilda, Bonnie J. Dorr, Walter L. Leite. (2026). Children's English Reading Story Generation via Supervised Fine-Tuning of Compact LLMs with Controllable Difficulty and Safety. https://arxiv.org/abs/2605.13709 doi:10.48550/arxiv.2605.13709

q2 · i? · design · r2

Qualitative analysis of toxicity outliers in the generated stories identified body-shaming and disgust-related language and unfriendly characters as contributing factors; such stories numbered five or fewer per 1,290-story dataset.

"We identified several potential factors contributing to the higher toxicity scores, such as the inclusion of language like "you're too fat," "this smells too stinky," and "the pig can gag," or the appearance of characters portrayed as unfriendly or potentially harmful, such as rats."

Discussion¶

Related Claims¶

• Rewarded SFT produces the most stable outputs among the SFT designs, nearly free of hallucinations and story-irrelevant material — related
• Fine-tuned 8B LLMs generate K-2 stories with difficulty-related metric scores statistically significantly different from (mostly better than) zero-shot GPT-4o and Llama 3.3 70B — related
• Fine-tuned 8B models show suboptimal inter-story diversity, repeating names, animals, locations, and activities across narratives — related
• Among the three 8B LLMs, Llama 3 underperforms Apertus and Granite 3.3 on most evaluation metrics — related