Particle Sliders

Particle Sliders gives users adjustable control over how Meta’s Muse Glimmer behaves. Users can train small behavior adapters or use pretrained ones to make Glimmer sleepy, excited or change other aspects of its responses. Slider strengths can be adjusted and combined. The project includes training and sampling code, a local slider interface and a Hugging Face showcase with recorded examples across the slider catalog. Particle Sliders and ParticleGAN are open-source projects that existed before this competition and continued to be developed during it. The Glimmer-specific integration and adapter training were created for this competition. What makes this unique is that it was trained with a GAN and has a different lora formulation along with custom research that stabilizes the GAN. Normal slider formulation uses L2 loss (teacher-student)**2 which causes conditional averaging and breaks next token selection in autoregressive models in my testing. https://huggingface.co/ntc-ai/glimmer-particle-sliders full lora weights