Hanshu Rao, Guangzeng Han, Xiaolei Huang
View original ↗Create a 'Hardness-Aware' training wrapper that fits onto existing fine-tuning scripts. It should dynamically re-weight training samples based on model confusion during the run.
Suggested repo: hamrTune
"Stop over-training on easy data—adaptive meta-resampling for robust LLM fine-tuning."
Estimated effort: 45h