I designed an AI chatbot that generates diverse, memorable, and linguistically grounded mnemonic devices for learning English and Mandarin Chinese vocabulary.
The project used chain-of-thought distillation from a teacher model, DeepSeek-R1, to instill linguistic reasoning in a Gemma3-1B student model through supervised fine-tuning with QLoRA and unsloth.
I also implemented a Direct Preference Optimization (DPO) pipeline with trl, using 500 human- and LLM-annotated preference pairs scored on memorability, imageability, and expected retention.