2026-10-02 · America/Los_Angeles · 社区动态 · #14

Our generative music recommender showed no plateau with model size or training data in the ranges we tested

Our tech report on Sona, a generative music recommender in A/B testing at Yandex Music, has two scaling sweeps that might interest this sub. We trained three encoder-decoder sizes (20M, 130M and 260M transformer-core parameters) on 8,192-event histories. Training next-token loss at matched exposure kept dropping with each step up. That sweep reports training loss only. For the data sweep, we took the 130M backbone with 2k-event histories and grew the training window from 1 to 8 weeks, one pass each. Recall@10 on held-out requests went from 0.1798 to 0.2519 and was still rising from 4 to 8…

热度 45.0 / 100；排名与评分保留该期记录。