great solution to a real problem, fine-tuning failures are often data problems, but we rarely have visibility into which examples caused them. This kind of attribution could make debugging much faster
LameRs•34m ago
interesting approach, but why focus on attribution after training instead of during training?
abdullah-xyz•33m ago
Because many failures only become visible after the model changes, and understanding the cause afterward is where current workflows are weakest
rootHQW•40m ago