Tiger WangOPEN TO INTERNSHIPS
All work
ML2025

ML Research

Deep learning research — dual-objective GRPO for LLMs optimizing factuality and fairness, plus SFT pipelines with LoRA/QLoRA. Shipped a fully reproducible open-source toolkit.

FIG. 05ML RESEARCH05

01 — The research

Developed and trained deep learning models using PyTorch and TensorFlow, including custom training loops, GPU acceleration, mixed-precision training, and modular model components for rapid experimentation through the use of RunPod.

02 — GRPO

Led a new research line on dual-objective GRPO for large language models, optimizing for both factuality (hallucination reduction) and demographic fairness through end-to-end training, judge prompts, and statistical testing.

03 — The toolkit

Shipped an end-to-end, fully reproducible toolkit — data cleaning, training scripts, evaluation plots — open-sourced with a runnable Colab for quick replication. Implemented SFT pipelines with LoRA/QLoRA adapters and experiment tracking with W&B/MLflow.

SPECIFICATION
ROLEML Researcher
YEAR2025
STACK
PyTorchTensorFlowLoRAW&BRunPod
FocusLLM Training
MethodGRPO + SFT
OutputOpen-source
NEXT PROJECTIMC Prosperity 4