Endless Lup: Training Creative LLMs

This is an intermediate version of this blog post. A complete version with detailed explanations, examples, and visualizations is coming soon.

Shoutout to Chase Harmon and Jon Ebataleye for their assistance as UROPs on this project.

In the winter of 2024, I began working on a 24/7 radio of Grammy-award winning rapper Lupe Fiasco (press here).

The main challenge was training a language model to emulate Lupe's distinctive style. Here's some techniques we found most helpful in training creative LLMs:

Dataset Curation & Evaluation

• Aggressively filtered dataset to remove low-quality and "low-creativity" samples
• Balanced corpus of lyrics, interviews, and social media content
• Grew non-Lupe dataset using LLM-measured similarity to Lupe's lyrics

Multi-Task Training Architecture

• Built LLM's creative skills incrementally through auxiliary tasks
• Still optimized for outputs to always be in Lupe's style to nail style transfer

Pretrained Model Optimization

• Used targeted fine-tuning to preserve capabilities while removing all "LLM slop"
• Balanced creativity with style fidelity in additional loss functions

Advanced Prompt Engineering

• Built prompt templates for stylistic nuances
• Used context-aware prompting for thematic consistency
• Created feedback loop for prompt refinement
• Controlled creativity level in outputs

Evaluation Framework

• Measured both technical and artistic quality
• Used human-in-the-loop evaluation
• Used A/B testing for comparing outputs
• Measured style transfer effectiveness

FULL BLOG POST COMING SOON

This is an intermediate version of the blog post about Endless Lup. A complete blog post with more details of our approach, examples, and insights coming soon!