IRI

IRI is a Korean voice conversation sandbox for children aged 4 to 10. I fine-tuned Kanana 2 3B with QLoRA for age-appropriate responses, served the model on Runpod GPUs, and built a speech-to-text and text-to-speech interface. Input and output checks in the API layer apply to both model responses and the guarded fallback route.

Scenario evaluations exposed safety failures that validation loss alone did not reveal. I used those cases to expand corrective training data and repeat the evaluation cycle. The latest adapter was checked after a fresh reload, while a final independent safety evaluation remains pending.