Search papers, labs, and topics across Lattice.
Data Provenance Initiative, University of Southern California
1
0
3
Don't count on unembedding matrix geometry to predict language model performance—it's more a reflection of training hyperparameters than inherent capabilities.
