Gradian

Gradian이란?

Training-data attribution for LoRA fine-tuned LLMs. Gradian scores every training example against the capability you lost, then runs a battery of dataset and config checks for the failures influence math misses.