Type: url
Original: https://habr.com/ru/companies/yandex/articles/841436/
Status: partial
Summary
A September 2024 post on Yandex’s Habr blog describing how an ML laboratory in Yandex Infrastructure built the inline code suggestions of Yandex Code Assistant, from statement-level prediction and fine-tuning on internal code to a split CPU/GPU backend. Most of it concerns evaluation: retention as the business metric, an offline unit-test metric, online acceptance rate, and finally a composite “developer happiness” metric that the team found to track retention better than acceptance rate.