sparse-mixture

Introduces PRISM, a prototype language model architecture that uses sparse, non-negative mixtures of learned prototypes for interpretable sequence modeling, achieving competitive performance and enabling fast training data attribution and model editing without fine-tuning.