AmelieSchreiber/toricgt-curated-splits
Viewer • Updated • 5.79M • 2.13k
A retraining of a TokenGT, embedding space GFlowNets GoT reasoning 4x4M-Soft-MoE model with tropical ring attention and geometric algebra constraints
Note ToricGT 17M-class <20K steps checkpoint repo with TokenGT graph tokens, tropical ring attention, default Soft-MoE, and embedding-space GFlowNet head to optimize for low BPB (bits-per-byte)