Unembedding Matrix

An unembedding matrix is the final learned weight matrix in a large language model (LLM) that maps the model’s output vectors back into vocabulary token probabilities.