← Back to Architecture

SWA

Architecture
Used in
11 PRs
Best BPB
0.0180
Avg BPB
0.8987

Hyperparameters Across PRs

pr_numberparameters
484
610{"frequency_steps":50}
1030
1056
1512{"window_size":256,"full_attention_layers":5}
1544{"window":512,"heads":8,"kv_heads":4,"shared":true}
1545{"window":512,"heads":8,"kv_heads":4,"shared_weights":true}
1553{"layers":2,"window":512,"heads":8,"kv_heads":4,"shared_weights":true}
1632{"layers":2,"window":512,"heads":8,"kv_heads":4}
1672{"window":512,"qk_gain":5}
1749{"layers":2,"shared_weights":true,"window":512}