The updated models are 0.4B, 1.5B, and 3B. Other parameter sizes will be updated. RWKV-Lite has two types of SVD
- Simple SVD with continual pretraining (-svd.pth): its model is inherited by the official checkpoints (rwkv-5) and modified with the standard singular value decomposition.
- Enhanced SVD with normal pretraining (-enhanced-svd.pth): its model is pretrained from scratch and applied with the enhanced SVD for model capacity.