The Electronic DJ Model (EDM) is an transformer model trained from scratch that generates a DJ playlist track-by-track curated to your initial track selection. It learned from nearly 5,000 real DJ sets in the DJ Mix Dataset.

The Idea: A DJ mix is an ordered list of music tracks, similar to how a sentence is an ordered list of words. We developed a novel tokenizer to handle the 63,040 unique tracks in the DJ Mix Dataset.

Limitations: With only ~96K tracks of training data, the big models memorized mixes. A small ~8M-parameter model did best: it learned patterns instead of copying them.

Continuous Improvement: Press thumbs-up to give positive feedback. Press thumbs-down to replace with a new track. Preferences are anonymously captured for DPO improvement to the model

Despite these limitations, I have enjoyed listening to the generated DJ sets. How was your personalized DJ set?

Read the technical blog here: https://anmokulabs.com/blog/electronic-dj-model.html