LoRA fine-tune of nvidia/Llama-3.1-Nemotron-Nano-4B-v1.1 for Plushie Tales, a fully local bilingual
children's story generator built for the Build Small hackathon (June 2026).
Trained on ~300 synthetic bilingual (English + interleaved Spanish) children's
stories distilled from NVIDIA Nemotron 3 Nano 30B-A3B (served via Ollama) and filtered through the app's quality gates:
Flesch-Kincaid readability bands per age (3-10), bilingual format validation, and a
child-safety blocklist. The goal is first-try compliance with the app's strict output
format, replacing a 7B teacher with a 4B student.
Runs entirely on-device. Part of a pipeline with BLIP-2 (vision), Kokoro + Piper (TTS),
and ChromaDB RAG — total stack well under the hackathon's 32B parameter cap.
