Good LLMs need BAD Data The Shocking Truth by HARVARD
AI Summary
In this video, titled “Good LLMs need BAD Data: The Shocking Truth by HARVARD”, presenter Howard delves into the relationship between pre-training data and the performance of language models. He argues against the traditional emphasis on using only clean data for training, suggesting that including a certain percentage of ‘bad’ or toxic data can improve an LLM’s alignment and robustness. Howard references a Harvard study by Kenneth Li et. al., which indicates that a mixture of clean and toxic data helps the model better recognize and handle undesirable outputs during post-training processes like supervision and reinforcement learning. By strategically incorporating bad data during pre-training, he highlights the potential for creating AI systems that are easier to control and more effectively aligned with human expectations. The video emphasizes the importance of understanding how data shapes model behavior and advocates for a holistic approach in the development of LLMs.