Tip Of AI
Discover what you can do with AI
Electronic DJ Model is a transformer based music recommendation system that generates DJ playlists one track at a time. Users choose an initial song and the model produces a continuous mix with related tracks. The system was trained from scratch using 5040 real DJ sets collected in the DJ Mix Dataset. It was designed to understand the structure and progression of DJ mixes, treating an ordered sequence of tracks similarly to a sentence made from words. To process the music catalog, the developers created a specialized tokenizer capable of representing 63040 unique tracks. The available training data included approximately 96000 tracks. Testing showed that larger models tended to memorize existing mixes, while a smaller model with approximately 8 million parameters performed better by learning general patterns instead of copying examples. The interface allows users to start a mix, request another track similar to the current selection, view the generated tracklist, and provide positive or negative feedback. Anonymous preferences are collected to support Direct Preference Optimization and future model improvements. The service also provides access to a technical blog post explaining its development.