Machine learning researcher in Canada. M.Sc. in Computer Science (thesis-based) from the University of Alberta, where I worked with Dr. Lili Mou at Amii on natural language processing.

My work sits at the intersection of reinforcement learning and large language models: how to turn an LLM's judgement into a training signal, how to distill that knowledge into smaller models, and how to evaluate text generation honestly. I care about ideas that hold up under careful experiments and about explaining them clearly.

Contact: donghengli9@gmail.com

LLMR: Knowledge Distillation with a Large Language Model-Induced Reward Dongheng Li, Yongchang Hao, Lili Mou. LREC-COLING 2024. Paper · Code · Lab

Instead of training a student model to imitate a teacher's outputs token by token, LLMR asks a large language model to judge the student's generations and uses that judgement as a reward in a reinforcement learning objective. The student learns from the teacher's preferences rather than its exact wording, which reduces exposure bias and improves BLEU/ROUGE by up to 36% over distillation baselines on dialogue and summarization tasks.

M.Sc. thesis (University of Alberta, 2025): LLMR: Knowledge Distillation with a Large Language Model-Induced Reward. Committee: Lili Mou, Di Niu, Davood Rafiei.

Interests: reward modeling and RL fine-tuning of language models, knowledge distillation, sequence-level training objectives, evaluation of text generation.

Tools I maintain: Seq2seq_ddp_kd, a training framework for loading multiple models under DistributedDataParallel for distillation experiments. CHAT-API, a multi-tenant assistant backend that puts a streaming chat widget on a website with one script tag (demo); in production on two commercial sites.

Teaching Assistant, CMPUT 466/566 Machine Learning, University of Alberta (2023–2024). Wrote and maintained the course's Jupyter/Colab notebooks, led labs and office hours, and graded with written feedback.

ml-teaching-notebooks: beginner-friendly Colab notebooks, one ML idea each, with leakage-free feature engineering, honest evaluation, and exercises.

• ShiftMaster: scheduling and payroll SaaS for small hourly-workforce businesses, 300+ monthly active users. Founder and sole developer. A per-company XGBoost model learns each business's past scheduling habits and drafts the week; the payroll engine handles Canadian statutory deductions and is covered by 60 unit tests with independently computed expectations. React, Supabase (Postgres with row-level security), Stripe, FastAPI, Google Cloud Run. Source private.
• A.H. Studio: commercial website for a children's art education studio, with a bilingual RAG chat assistant that answers enrolment questions around the clock (Pinecone retrieval, SSE streaming, Gemini with fallback to other providers). Since launch the studio's enrolment is up more than 30%. Plain-language guides let the owner update content and redeploy without a developer. React, FastAPI, LangChain, Google Cloud Run. Website source private; the assistant runs on CHAT-API above.

PyTorch · Hugging Face · reinforcement learning · XGBoost · scikit-learn · FastAPI · React · Google Cloud