SHA-RNN (4 layers, h=1024, attention head per layer)

Reported on 1 benchmark across 1 task · 1 paper

Note: results are matched by exact model name. Different papers may use the same name for different model variants.