BenChaliah/Superposition-Transformer

a novel architecture that leverages Autoencoders to superimpose the hidden representations of a base model and a fine-tuned model within a shared parameter space. Using B-spline-based blending coefficients and autoencoders that adaptively reconstruct the original hidden states based on the input data distribution.

语言：Jupyter Notebook · GitHub 星标：78 · Fork：3

话题：artificial-intelligence · autoencoders · deep-learning · language-model · llm · llm-training

在 GitHub 上查看原仓库