Models fine-tuned by minimizing KL between each layer and the final unembedding matrix