Spotify fine-tunes Llama to explain recommendations and write AI DJ commentary

The problem

Spotify's listeners decide what to play from cover art and what they already know. Spotify wanted to add context that works like a friend's tip, so people would try new artists, podcasts and audiobooks. Spotify saw large language models as a way to scale this kind of personal narrative.

What they built

Spotify used Meta's Llama models, whose files are published openly so a company can adapt them, to write two things. The first is a short explanation of why an item is recommended. The second is spoken commentary for its AI DJ, a personalised guide that picks songs and talks about them, in English and Spanish. Fine-tuning means training a model further on a company's own examples so it fits a specific job.

How they did it

Spotify first tried Llama with plain prompts and found the output needed tighter control. Expert editors wrote model examples and kept giving feedback on errors in artist names, tone and facts. The team added instruction tuning, prompt work and tests that try to provoke bad output. For the DJ, Spotify compared outside and open models and found that fine-tuned smaller Llama models gave commentary on a par with leading models at lower cost and delay. It also served models with the vLLM inference engine. Separately, after Llama 3.1 8B performed competitively in tests, it adapted Llama to 10 internal tasks.

Results

Spotify reports that in online tests, users were in some cases up to four times as likely to click on a recommendation that came with an explanation, especially for niche content. Meta's post describes it as users being four times as likely to engage with content from a creator when told why it was recommended. Adapting Llama to Spotify's tasks improved results by up to 14% over the stock model, with little loss in general ability. For the DJ, Spotify says listeners were more willing to play songs they might otherwise skip.