Full talk · edited transcript

Clean Labels for ML

Anastasia Bezzubtseva27 May 2023

Anastasiya Bezzyubtseva, Head of Voice Analytics Group, Yandex. Yandex Track at.

It is believed that a significant portion of the success of ChatGPT is due to it being trained on very clean data. The talk discusses how clean data and labels affect quality in other applications, not only in large language models. The speaker has worked at Yandex for almost 10 years: metrics, data science, model training, articles, research; initially anti-spam, then search ranking, a bit of Toloka, now Alice and the NLP department, where she leads the voice recognition and synthesis analytics group.

Plan: what kinds of labels exist and how dirty they can be; tips for cleaning labels; practical stories about how label cleaning improved model quality.

Continue reading — it's free

Google or a one-time email code. No password.