vec2vec-vsp-ablation-cos-drop

Summary

The ablation results show that the VSP loss is doing essential work in the model: taking it out causes cosine similarity to nearly halve, one key metric to zero out, and the ranking score to balloon to over 4,000. In practical terms, this means the system cannot drop that loss term from training without severely degrading the quality of its learned representations, so any future changes to the training pipeline must account for its role.