Skip to content

QLoRA-based SFT pipeline for phoneme-controlled children's story generation with three 8B LLMs and four experimental designs

Design · All designs
Evidence · 2 claims (2 for) · 1 study (1 design), q2 · 1 of 1 report an effect size · 2 claims rest on one study

Description

The article builds a pipeline that fine-tunes Llama 3 8B, Apertus 8B-instruct-2509, and Granite 3.3 8b-instruct using QLoRA (4-bit NF4 quantization with LoRA adapters, rank 32, alpha 64) on input-target pairs pairing curriculum lesson phonemes with 2,580 stories generated zero-shot by GPT-4o and Llama 3.3 70B from the 129-lesson UFLI K-2 curriculum. Four designs are compared: baseline SFT, "Good Stories" quality filtering (996 stories selected), Rewarded SFT with reward-weighted loss, and SFT with simulated children's reading errors generated by GPT-OSS-120B. Evaluation uses Spache readability, LM-PPL, coherence, syntactic complexity, toxicity, and Self-BLEU.

Design Implications

Context

Requirements

  • Training data derived from an expert-designed K-2 phonics curriculum specifying target phonemes per lesson
  • Sufficient stories per lesson as SFT targets (2,580 stories, 50-350 words each)

Constraints

  • Computational resources limited to three L4 GPUs, preventing some SFT strategies and constraining hyperparameters such as training epochs
  • Sample size was insufficient for a full RLHF pipeline
  • Only a single expert-designed curriculum and one language were used

Target Learners

  • Children in kindergarten through 2nd grade learning English reading

Learning Goals

  • Phoneme-targeted reading practice
  • Oral reading fluency through engaging level-appropriate stories

Claims

Examples

-

Key Sources

  • 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/2605.13709v1
Page metadata
Field Value
type design
id ufli-sft-story-generation-pipeline
title QLoRA-based SFT pipeline for phoneme-controlled children's story generation with three 8B LLMs and four experimental designs
description The article builds a pipeline that fine-tunes Llama 3 8B, Apertus 8B-instruct-2509, and Granite 3.3 8b-instruct using QLoRA (4-bit NF4 quantization with LoRA adapters, rank 32, alpha 64) on input-target pairs pairing...
status review
generated by process:settle-candidates on 2026-10-11
trust unverified — no one has confirmed this page's content
sources 1 — qian-shen-2026

Frontmatter is the page's machine-readable record — see the Schema & Guide.